Detection and identification of foreign bodies in conditioned steak based on ultrasound imaging

Abstract

Conditioned steak is easily contaminated by foreign bodies, such as iron sheets, glass, and crush bones in the manufacturing processes, posing hidden safety hazards to consumers. In this study, the feasibility of using ultrasonic imaging to detect and identify foreign bodies in conditioned steaks was investigated. Firstly, the ultrasonic imaging data of foreign bodies was collected. Four discriminant models among them linear discriminant analysis (LDA), and extreme learning machine (ELM) were established, and based on the texture values of the smallest circumscribed rectangular area of the foreign bodies, the type was identified. The foreign bodies were then extracted by gray–level co–occurrence matrix (GLCM). The detection rate of foreign bodies was 97.78 %, meanwhile ELM showed the highest accuracy of recognition rate of 76.67 %. The results showed that ultrasound imaging technology could be used to detect foreign bodies in the conditioned steak and to identify the type of foreign body via pattern recognition.

Introduction

Conditioned steak is a kind of treated meat with advanced and recent techniques to witness a delicious taste, rich nutrition, convenient consumption, and low price as per the consumers’ needs (Wang et al., 2020). As market demand has increased, the manufacture of conditioned steaks has gradually transformed from traditional manual production to mechanized industrial production. In a large–scale production using huge machines, steaks may be mixed with foreign metal bodies due to wear of mechanical parts. Moreover, the steaks may be contaminated with glass, broken bones, and other foreign bodies during the slaughtering, rolling, and picking processes, which carries serious health risks and negatively affects the satisfaction of consumers. The quality of food products is not only a matter of substandard quality, shelf life and counterfeiting, but also the safety and convenience of food. The presence of foreign bodies in food means that the quality of the food is not up to scratch, which not only jeopardizes consumer safety but also undermines public trust in the food brand. Therefore, it is necessary to have a rapid and effective detection method to keep a highly efficient primary quality control.

Previous studies on meat products had chiefly focused on quality analysis but foreign body detection was not thoroughly studied. Common methods for detecting foreign bodies in food have certain limitations, för instance magnetic metal detection (Fortaleza et al., 2018) can only detect foreign metal objects, where optical detection (Srivastava et al., 2020; Shi et al., 2023; Li et al., 2020; Ravikanth et al., 2016) can only detect the foreign bodies on shallow surfaces, whereas X–ray (Nielsen et al., 2013; Kim, 2019; Voss et al., 2021) has several disadvantages, such as the high cost and high power usage, especially when applying nuclear magnetic resonance (Koizumi et al., 2010; Suchanek et al., 2017).

Ultrasonic waves, with frequencies exceeding 20 kHz (Zhang et al., 2021), are a form of mechanical wave that propagates through elastic mediums. In addition to adhering to fundamental acoustic principles during propagation, they exhibit favorable characteristics such as excellent directionality, concentrated energy, and long transmission distances (Cheng et al., 2014; Jiang et al., 2021). Regarding the energy production, ultrasound can be divided into high–energy which would produce thermal and cavitation effects during propagating in the medium and be commonly used in cleaning (Liu et al., 2014) and extraction (Xu et al., 2020; Wang et al, 2021; Wang et al., 2021). Low–energy ultrasound is often used in ranging (Collado–Lara et al., 2022) and detection (Li et al., 2021), as it does not change the structural properties of substance. Hæggström et al. (2001) used 5 MHz ultrasonic waves to detect the signals of samples with and without foreign matter, and it was proven effective in identifying common foreign matter in liquid–semi–solid foods, such as jam, butter, and cheese. Leemans et al. (2009) analyzed the reflected signals of the ultrasound to distinguish the types of foreign bodies in the cheese. Ultrasonic detection of foreign bodies has a wide range of usage, low cost, with no pollution side–effects, and able to provide a certain depth of inspection.

In this line and compared with general ultrasonic testing, the reflective ultrasonic imaging is more intuitive and sensitive. This method also allows the collection of more information, such as the location and size of the foreign body. Ultrasound imaging technology is widely used in industrial inspection (Song and Ni, 2018; Chaitanya and Kumar, 2021) and medical diagnosis (Guo et al., 2018; Stone and Quiroz, 2016) based on its fast, non–destructive, highly sensitive performance, etc. Nonetheless, it is rarely used for foreign body detection in food. Cho et al. (2003) used a transmissive air–coupled ultrasonic imaging system to detect foreign bodies including glass and metals in chicken and cheese. According to the speed and attenuation coefficient of ultrasound propagation, the imaging study used image processing methods such as interpolation to optimize the ultrasound image deformed by diffraction and refraction to achieve foreign body detection. Ultrasound imaging is highly adaptable and easy to be automated. Notwithstanding, ultrasound images showed low resolution, blurred contours of foreign bodies which hinders the identification of deep foreign bodies in food by transmission imaging. However, it is possible to detect the spatial position of foreign bodies in food by reflective ultrasound imaging, which shows great potential in the detection strategies (Payel et al., 2016).

Therefore, this study explored the feasibility of using ultrasonic imaging technology to detect and identify foreign bodies in conditioned steaks. Firstly, acoustic parameters of the conditioned steak were measured. Ultrasonic detection signal is based on the difference in the density of objects. Fortunately, steaks and foreign bodies exhibit different density and acoustic impedance, facilitating the collection of ultrasound images. The reflection waveforms and ultrasound images of foreign bodies were analyzed. Penultimately, the foreign body region was segmented by image processing, and the detection rate was calculated. Finally, the texture feature values of foreign bodies ultrasonic images were extracted, and the identification model was established by the method of pattern recognition, which is expected to provide a reference for applying ultrasonic imaging technology to detect and recognize foreign bodies in meat products.

Materials and Methods

Test materials　The raw material of the steak used in the experiment was sirloin (produced in Australia), which was purchased from local Metro supermarket in Zhenjiang. The auxiliary materials used in the experiment (edible salt, white sugar, and spicy materials) were purchased from Zhenjiang Metro supermarket. The food additives used in the test were sodium bicarbonate and complex phosphates (sodium tripolyphosphate and sodium hexametaphosphate), which were purchased from Henan Qianzhi Trading Co., Ltd.

Operation points　The raw meat was trimmed, the fascia and blood clots were removed. Then, the meat was cut into 12 mm thickness for each piece. Each 1 kg raw meat was mixed with 15 g salt, 5 g white sugar, 20 g spices, 3 g sodium bicarbonate, 3 g compound phosphate, and 150 mL of water which was dissolved as a salted liquid. The raw meat and salted liquid were placed in a vacuum pickling machine (KA–6189, Shenzhen, China), kneading for 1h. Finally, the steaks were discharged into trays for sealed packaging and stored at 4 °C for no more than one day.

Ultrasound imaging system　The ultrasonic micro–imaging system used in the experiment was independently developed by the food non–destructive testing laboratory of Jiangsu University. The schematic and physical map of ultrasonic imaging system were shown in Fig. 1. The hardware part of the system included the UTEX 320 ultrasonic signal transmitter/receiver (UTEX Scientific Instruments Inc., Canada), a 20 MHz point focusing ultrasonic transducer (Olympus Corporation Inc., Japan) with a diameter of 10 mm and a focal length of 25.4 mm, a high–speed A/D data acquisition card (Agilent Technologies Co., Ltd., America), a three–axis precision linear motor scanning mechanism (Chuang Feng Seiko Co., Ltd., China), a computer, sample tanks.

Fig. 1

Ultrasonic imaging system. (a) Schematic diagram; (b) Physical diagram: (1) Ultrasonic signal transmitter/receiver; (2) Data acquisition card; (3) Computer; (4) Motion control card; (5) Main control circuit; (6) Three-axis precision motor scanning mechanism; (7) Ultrasonic transducer; (8) Sample; (9) Water platform; (10) Software interface of ultrasonic microscopic imaging system.

Measurement of acoustic parameters of conditioned steak　Calculating the depth of foreign bodies in the steak required measuring the propagation speed of the ultrasound. The gray value of ultrasonic imaging pixel points depended on the ultrasonic reflection echo intensity which was based on the difference in the acoustic impedance on the ultrasonic propagation path. Therefore, the steak acoustic impedance values need to be measured and compared with different foreign bodies to explain the differences in detection results of foreign bodies with varied acoustic impedances in steak. The acoustic characteristic parameters of the steak, sound velocity and acoustic impedance were measured from the above. The schematic diagram of steaks acoustic parameters measurement was shown in Fig. 2. Since the ultrasonic propagation was susceptible to temperature, the measurement was carried out at a temperature of 20 °C. The transducer was 11.2 mm away from the steak and scanned at a speed of 5 mm/s. Each measurement took 3 min and the experimental data were expressed as mean of 5 replicates. ENVI software (ITT Visual Information Solutions, Boulder, CO, USA) was used to derive the ultrasonic image information of the sample, and selected the target area with the unit of 200 × 200 pixels in the center of the steak sample.

Fig. 2

Schematic drawing of the steak acoustic parameters measurement. (d), thickness of steak, m; (t₁), time for the ultrasonic wave to return from the upper surface of the steak, s; (t₂), time for the ultrasonic wave to return from the bottom of the steak, s; (Z), acoustic Impedance, N·s/m³; (F₁), the echo of the ultrasound on the steak surface; (F₂), the echo of the ultrasound on the underside of the steak.

The ultrasonic speed in the steak could be calculated according to the thickness of the steak and the ultrasound propagation time in the steak. The calculation was formulized as follows:

  

d—Thickness of steak, m;

t₁—Time for the ultrasonic wave to return from the upper surface of the steak, s;

t₂—Time for the ultrasonic wave to return from the bottom of the steak, s.

  

Z—Acoustic Impedance, N·s/m³;

ρ—Density of steak, kg/m³;

ν—The ultrasonic speed in the steak, m/s.

Foreign body implantation　This study was an improvement based on previous literature study (Joanna et al., 2016), and three foreign bodies (size of 3 × 3 × 1 mm and 6 × 6 × 1 mm), iron sheets, glass, and crush bones were selected. Due to the 12 mm thickness of the steak, the embedding depth was selected to be 3 mm, 6 mm, and 9 mm. The steak prepared in Material and Methods was cut into small pieces for foreign body embedding detection. The sample and foreign bodies were shown in Fig. 3. The foreign body was embedded from one side of the steak. The samples were tested after being placed in a refrigerator at 4°C for 2 h to reduce the impact of the embedding process on the detection. Each foreign body was scanned 10 times at different depths, and a total of 180 ultrasound imaging data were obtained.

Fig. 3

Physical image of samples and foreign bodies.

Ultrasonic image acquisition and processing of foreign bodies　The main parameters of the ultrasound imaging scanning system included pulse voltage (determining the strength of the ultrasound signal), pulse width (pulsing duration), pulse repetition frequency (number of ultrasound transmissions per unit time), gain (magnification of ultrasound signal), resolution (resolvable size of the smallest target) and the scanning speed. Different combinations of ultrasonic image acquisition parameters were set to perform multiple scanning tests. The apparent resolution of the images obtained under different parameters was evaluated by Tenengrad gradient function to determine the best parameters for preparing ultrasonic steak image acquisition.

The Sobel operator was employed to separately extract horizontal and vertical gradient values. The definition of image sharpness based on the Tenengrad gradient function was as follows:

  

  

T denoted the specified edge detection threshold. Gx and Gy represented the convolutions of the Sobel horizontal and vertical edge detection operators at pixel point (x,y), respectively. The Sobel operator template was employed for edge detection:

  

The higher the return value of the Tenengrad function, the clearer the image. By comparing the return values of different parameter combinations, the optimal parameters for ultrasound image acquisition were determined. The optimal parameters were 250 V pulse voltage, 25 ns pulse width, 800 Hz pulse repetition frequency, 50 dB gain, 0.1 mm resolution, and 3 mm/s scanning speed.

Foreign body ultrasonic image processing　The overall flows of image processing consisted of image filtering, image enhancement, image segmentation, and morphological processing. Taking an ultrasound image of an iron sheet with an embedding depth of 6 mm and a size of 3 × 3 mm as an example, the process of segmenting foreign bodies from the image was followed. The gray values of noise in the figure were small, and most of them were pixels, which belonged to salt and pepper noise. The median filter had a good suppression impact on pepper and salt noise, so a 3 × 3 template was used to apply the median filter to the image. The gray value of the filtered image was linearly stretched the gray value of image to the interval of 0–255. The principle is as follows:

f(x,y) is the pixel value at the coordinate of (x, y), and its minimum grey level A and maximum grey level B be defined as:

  

Linearly mapping A and B to 0 and 255 respectively, the final image g(x,y) is obtained as:

  

Image segmentation was an essential step in image processing and referred to the technical process of separating the target area from the image. The maximum inter–class variance method, referred to as the Otsu’s method, could adaptively determine the threshold of segmentation. In view of the appearance of voids in some foreign body images, holes were filled. There were some isolated noise points in the edge contour region of the binary image. So, the opening operation and then the closing operation were used in image processing. The image contour was smoother without changing the area of the target area as much as possible. Finally, the minimum circumscribed rectangle of the boundary of a foreign body was made to extract the texture feature value of the next step.

Extraction method of image feature value　In ultrasound imaging, two–dimensional pseudo–color images were used to characterize the response distribution of the measured object to ultrasonic signals. The acoustic characteristics of the detected object could be quantified and analyzed by the extracted image feature values. The properties of the object itself was reflected and two different objects (or two images) were distinguished by texture features. Texture was a change in the gray level or color of an image pixel, and this change was spatially statistically related (Yelampalli et al., 2018; Ropelewska and Szwejda, 2018), which was an essential means of image analysis. Gray–Level Co–occurrence Matrix (GLCM) was a classic and widely used texture feature extraction method based on statistical laws, which reflected the second–order texture information of the spatial distribution of gray image levels, that can be defined as the joint probability density (P) between the gray image levels, and the formula for P is shown as follows:

  

d—Pixel pitch;

θ—Relative direction, generally take four directions of 0°, 45°, 90° and 135°; (x,y)—Pixel coordinates;

N—Total number of pixels in the image.

Calculating the gray co–occurrence matrix from 0°, 45°, 90°, and 135° and extract the downward angular second moment (ASM) (Stankovic et al. 2016), contrast (CON), correlation (COR), and homogeneity (HOM)), and calculating the image entropy (ENT) (Kumar et al. 2020), anisotropy (ANI), average (AVG) and variance (VAR) (Ata et al. 2019), a total of 20 texture feature variables were achieved.

Pattern recognition and classification method　Identifying the type of foreign body enables an analysis of its origin, facilitating the implementation of preventive and control measures to mitigate the recurrence of similar incidents. This holds significant implications for ensuring food safety. After the foreign body image was segmented, the foreign body image features could be extracted to establish a discriminant model to classify the foreign body type. In this paper, we used GLCM to derive the texture feature values of the smallest circumscribed rectangular area of the foreign body boundary from all 180 ultrasound images, split the data randomly, take two–thirds of the samples as the training set, and one–third of the samples as the test set. Due to the overlapping and redundant information between the image texture variables collected by GLCM, the predictive ability of the identification model will decrease. Principal component analysis (PCA) was used to reduce the dimensionality of the collected image texture variables to obtain all the information that characterizes the characteristics of the sample image. Using raw data to build the model leads to overfitting. Because of the large amount of original data, PCA could be used to reduce the dimensionality of the data. Extracting feature variables from the original image to build a model could improve the recognition effect of the model. The guiding principle for principal component selection was that the cumulative variance contribution should exceed 80 % and the recognition rate should achieve the highest. Simultaneously, in order to prevent model overfitting, the number of principal components was typically limited to no more than ten. Based on the extracted image texture feature values of foreign bodies, different classification algorithms were used to establish a foreign body type identification model. The discriminant model could be divided into linear and non–linear classification algorithms. The first few principal component variables of the image features after PCA processing were used as independent variables, and the sample category was used as the dependent variable to establish linear discriminant analysis (LDA), K–nearest neighbor (KNN), back propagation artificial neural network (BP–ANN), and extreme learning machine (ELM).

LDA, also known as Fisher linear discrimination, was ensure the distance within classes as small as possible, thus achieving similar samples’ aggregation and separating samples between classes as far as possible (Abuzeina and Al-Anzi, 2018). It was more suitable for high–dimensional multi–class pattern recognition, which could reduce the dimension of the feature space to a certain extent and effectively extract classification information. LDA algorithm combined with principal component analysis was used to establish three foreign body recognition models. The input variables of the model were the texture feature data of sample ultrasonic images, which were mainly the first ten principal components determined by principal component analysis.

KNN has the advantages of no linear separation of calibration set samples and separate training process, handling multiple types of problems, and easy to implement. The test sample was determined to be the most frequent category among the K samples (Xie et al., 2017). If the K value is too small, the discrimination will be performed in a small neighborhood, and the model is easy to overfit. If the K value is too large, the selected results will be widely affected, causing irrelevant sample points to participate in the judgment and resulting in underfitting. KNN algorithm and principal component analysis were used to establish the discrimination models of three foreign bodies. The distance metric used in the k–NN model was Euclidean metric. The first 10 principal component numbers and 9 parameter K were selected to optimize KNN model, and the recognition rate of test set was used as the evaluation basis of optimization results.

BP–ANN is a multi–layer feed–forward neural network trained by an error back propagation algorithm consisting of an input layer, a hidden layer, and an output layer. BP–ANN process continues until the error was less than the set threshold or the number of iterations was reached. BP–ANN was commonly used to solve non–linear classification problems or deal with the non–linear model relationship between two related data matrices (Sun et al., 2021). The principal components from the principal component analysis were used as input to the model. A three–layer BP network structure (input layer, implicit layer, output layer) was chosen for the experiment and the network parameters of the model were optimized. The range of input layers was 1–10, based on the principle of principal component selection. Based on the pilot study, the range of hidden layers was selected to be 3–15 and the number of hidden layers was chosen based on the recognition rate of the samples. The number of neurons in the output layer was 1, that is the type of the foreign body in the conditioned steak. The number of neurons in the input layer was 7 (the number of principal component factors is 7), the number of neurons in the hidden layer was 9, and the number of neurons in the output layer was 1. The optimized BP network topology was 7–9–1. After trying, the target error was set as 10⁻⁸, the learning rate was set as 0.1, the momentum factor was set as 0.7, and the epochs was set as 1000.

ELM only needs to set the number of hidden nodes in the network, which overcomes traditional neural network problems with many training parameters, complicated iteration process, and easy to fall into a local minimum. It had the advantages of good generalization performance and fast learning rate (Junior and Backes, 2016; Li et al., 2021). The output layer of the ELM had no error nodes, and the error could be minimized by solving the output layer weights. The activation function of ELM was “sigmoidal”. The network structure of ELM model was determined by comparative tests of different network layers. The root mean square error (RMSE) and coefficient of determination (R²) were used as measurement criteria. When the network structure was 5, the error was minimal. The number of hidden nodes in each layer was determined by trial and error, which were 40, 30, 20, 10, and 5 respectively.

Results and Discussion

Foreign body waveform analysis　Fig. 4 (a), (b), (c), (d) are the ultrasonic echo signal diagram series of the four groups; conditioned steak without foreign body, steak embedded with iron sheet, glass, or a broken bone, respectively (embedding depth: 3 mm). Fig. 4 shows that the four images have complex irregular echo signals at 0–5 µs, because of the reflected echoes between the acoustic lens and the matching layer. The peaks around 8 µs and 24 µs were the echo signals from the upper and bottom surfaces of the steak, respectively. Unlike the blank control, the obvious echo of the waveform diagram with foreign bodies embedded at 3 mm in the vicinity of 13 µs could be explained by the echo signals of foreign bodies. There were also apparent differences between the echo signals of the tested foreign bodies. The echo amplitude of iron sheet was the largest, and the secondary wave was the least. The echo amplitude of the broken bone was the smallest, and the secondary wave was the most. According to the measurement explained in Material and Methods, the average propagation velocity of the ultrasonic waves in the conditioned steak was 1522.6 m/s, and thus the depth of the embedding place could be calculated from the waveform diagram. For instance, the surface and the foreign body signal appeared in Fig. 4 (b) at 8.8 µs and 12.6 µs, respectively. Considering the ultrasonic propagation round–trip distance, it could be estimated that the distance between the iron plate and the upper surface was 2.89 mm.

Fig. 4

Waveform figures of samples and foreign bodies. The green part is the position where the foreign body ultrasonic signal appears.

Foreign body ultrasound image analysis　Fig. 5 is an ultrasound scan of foreign bodies of different materials and sizes at different depths. The color of the pixel represented the intensity of the reflected echo at that position, and the right side was the scale of the signal intensity corresponding to the color. The background area in the figure had a weak signal, which was caused by system noise and incomplete uniformity inside the steak. The deeper the embedding depth, the smaller the size of the foreign body, the worse the ultrasonic accuracy of foreign body detection. The best accuracy of foreign body detection was the iron sheet, and the worst was the broken bone, as consistent with the waveform diagram. According to previous data (Birks et al., 1991), the acoustic impedances of iron sheet, glass, and broken bone at 20°C were approximately 45.4 × 10⁶ kg/(m²·s), 13.1 × 10⁶ kg/(m²·s) and 6.3 × 10⁶ kg/(m²·s). Taking into account the acoustic parameters mentioned in Material and Methods, the acoustic impedance of conditioned steak was 1.69 × 10⁶ kg/(m²·s). Ultrasonic signals are generated by reflection from the interface of different acoustic impedance media. The greater the difference in acoustic impedance, the stronger the reflected echo is, explaining the variation in the seen effect of foreign bodies of different nature. The difference between the acoustic impedance of the conditioned steak and the iron piece was the largest, and the difference from the bone was the smallest. Therefore, the best accuracy of foreign body detection was shown in the iron piece, and the worst was in the bone.

Fig. 5

Ultrasonic images of foreign bodies of different materials and sizes at different depths in samples.

Foreign body ultrasonic image processing　An ultrasound image of an iron sheet at an embedding depth of 6 mm and a size of 3 × 3 mm as shown in Fig. 6 was subjected to the process of segmentation. The gray areas were representing the values of noise in the figure which were small, and mostly were pixels, displaying the salt and pepper noise. A 3 × 3 template is used to apply the median filter to the image to remove the salt–and–pepper noise in the image. The gray value of the filtered image was linearly stretched to the interval of 0–255.

Fig. 6

Image pre-processing. (a) Original image; (b) Median filter; (c) Gray value linear transformation.

Image segmentation is an essential step in image processing and refers to the technical process of separating the target area from the image. The binarized image segmented by the Otsu’s method is shown in Fig. 7 (a). The appearance of voids in some foreign body images, was explained by the filling of the holes. In binary images, isolated noise points are often found within the edge and contour regions. To preserve the integrity of the target area as much as possible, morphological opening and closing operations are performed on the image. This process yields a smoother image contour without significantly altering the regions of interest. On the premise of maintaining the target area to the maximum level, the image contour becomes smoother. Finally, the minimum circumscribed rectangle of the boundary of a foreign body was made to extract the texture feature value of the next step. The above process is shown in Fig. 7.

Fig. 7

Image segmentation. (a) Threshold segmentation; (b) Morphological processing; (c) Minimum enclosing rectangle.

The image in Fig. 5 is subjected to the processing mentioned above and the result is shown in Fig. 8, and all foreign bodies were successfully detected. All the collected 180 foreign body images were automatically processed. Only 4 images were not detected, leading to a detection rate of 97.78 %. Taken together, the ultrasonic imaging technology had an excellent detection of the foreign bodies in steaks. Tang et al. (2022) presented a novel CNN based CFOD to detect the foreign bodies on the surface of a tobacco pack where the results achieve an mAP of 96.3 %. The possibility of rapid and accurate foreign body detection using hyperspectral imaging technique was confirmed by Kwak et al. (2021) showing a detection accuracy of the proposed algorithm that reached 95 %. Compared with these studies, the ultrasonic imaging technology used in this study proved effective with most likely the highest detection rate so far, and to the best of our knowledge.

Fig. 8

Threshold segmentation of the ultrasonic image.

The recognition rate results of each model are shown in Table 1, supporting the notion that the ELM model had the best recognition effect. The recognition rate of the training set and test set were 83.33 % and 76.67 %, respectively. The recognition effect of KNN model was the worst. The recognition effect of non–linear models (BP–ANN and ELM) was better than linear models (LDA and KNN). The reflection echo intensity of foreign bodies with deep embedding depth was lower. The recognition area and gray value in the corresponding ultrasound image were small, make it difficult for the identification. A simple linear classification model was not the optimum to provide a good classification effect. It also showed a non–linear relationship between the type of foreign body and the texture feature value of ultrasound image. Ultrasonic detection signal is based on the difference of object density where various objects have different density and acoustic impedance. Steak is a low–density object, while foreign matter is a high–density object. When ultrasonic detection is used, the detection rate of foreign bodies reaches 97.78 %. However, foreign bodies such as iron sheets, glass, and broken bones are all high–density objects, and the signal distinction of ultrasonic detection is not obvious. The recognition rate of the ELM established by the texture feature value is 76.67 %, approving a better classification effect. The comparisons of different technologies for foreign body detection are shown in Table 2. Studies indicated that ultrasound imaging combined with the ELM model could identify the type of foreign body in the conditioned steaks.

Table 1 Confusion matrix of classification results of the test set based on different identification models of the test set.

Discrimination model	Predictive value/Actual value	Iron sheet	Glass	Crush bone	Recognition rate of test set/%
LDA	Iron sheet	16	2	2
	Glass	1	13	6	68.33
	Crush bone	3	5	12
KNN	Iron sheet	14	1	5
	Glass	0	13	7	58.33
	Crush bone	6	6	8
BP-ANN	Iron sheet	16	3	1
	Glass	1	12	7	66.67
	Crush bone	3	5	12
ELM	Iron sheet	19	0	1
	Glass	0	14	6	76.67
	Crush bone	1	6	13

Table 2. Comparison of different technologies for foreign body detection.

Detection method	Detection object	Foreign body	Detection and recognition rate	Detect strengths and weaknesses	Reference
X-ray	minced beef, cultured sour cream	folded standard paper, cigarette butt, broken glass, afly, harlekin ladybug	Detection rate 59% Recognition rate /	The detection effect of paper and insect is good, but glass and other high absorption is poor. The detection of tiny similar fibrous structures in food substrates presents challenges in distinguishing between different foreign bodies.	Nielsen et al., 2013
Ultrasonic	cheese	a plastic cylindrical object having a diameter of 3 mm (a pen core)	Detection rate 90% Recognition rate /	Due to the high signal attenuation caused by the cheese texture, the acoustic impedance of the foreign object close to the cheese, and the small volume of the foreign object, the signal to noise ratio of the object echo is low, which has a great influence on the detection result.	Leemans et al., 2009
Ultrasonic	chicken breasts	bone fragments	Detection rate 74.8 % Recognition rate /	Amplitude ratio, not velocity, could successfully discriminate between uncut samples, cut samples, and cut samples with bone fragment with projected area from 6 mm2 to 16 mm2. However, intuitively, it appears that it would be more difficult to locate a smaller bone fragment at a higher vertical location.	Correia et al., 2007
Ultrasonic	cheese, marmalade	stone, piece of glass, woodsphere, wood cube, plasticsphere, bone sphere, steel sphere	Detection rate 71.43 % Recognition rate /	It can be used to detect foreign bodies in homogeneous products at depths of 20 – 75 mm. Due to the poor S/N ratio, the detection depth of heterogeneous products is limited to 50 mm, which makes it difficult to detect foreign bodies in heterogeneous products.	Hæggström et al., 2001
Dedicated Micro-Magnetic Resonance Imaging	young apple fruit	peach fruit moth, carposina sasakii	Detection rate/Recognition rate/	The 1 T dedicated MRI device is small, easy to operate and maintenance-free; it can be used for a specified subject without restriction of operation times by placing it in a lightly air-conditioned ordinary research room. Since the measurement cell was just 30 mm in diameter, the apparatus could not deal with harvested apple fruits.	Koizumi, et al., 2010
Ultrasound Imaging	cheese, poultry	glass, steel rod	Detection rate/Recognition rate/	The minimum 3 × 3 mm2 foreign fragments in cheese and poultry and 1.5-mm-dia cylindrical objects in the food materials could be clearly detected. Further study is needed on NCU imaging of food materials of uneven or irregular shapes for calibration and use in online detection under actual production setting.	Cho et al., 2003
Ultrasound imaging	Conditioning steak	ron sheets, glass, and crush bones	Detection rate 97.78 %, Recognition rate 76.67 %	It is more intuitive and convenient, and the detection rate of foreign body locating by plane scanning is high, and the probability of misdetermination is small, The recognition rate of the three foreign body identification models is good, which can provide some guidance for the backward tracing of foreign body detection, However, due to the limitations of ultrasound, the sample needs to be placed in water, which needs to be further studied and solved.	This study

Conclusions

Some acoustic parameters of the conditioned steak were measured for the first time in this article. Ultrasound images of the conditioned steak embedded with foreign bodies were collected, and then image processing was used to segment the foreign body regions. The results showed that the average propagation velocity of ultrasonic waves in the conditioned steak was 1522.6 m/s, and the acoustic impedance of the conditioned steak was 1.69 × 10⁶ kg/(m²·s), which could be used to calculate the depth of foreign bodies and recognize the possible different types of foreign bodies. Through a series of image processing, foreign bodies could be successfully separated from the background image, and the total detection rate of all foreign body ultrasound images was 97.78 %. Among the four identification models (LDA, KNN, BP–ANN, and ELM), which were built to distinguish between iron, glass and broken bone, ELM showed the best result (the recognition rate of the test set was 76.67 %).

Acknowledgements　The authors gratefully acknowledge the financial support provided by National Key Research and Development Program of China (2016YFD0401104), Open Project Program of National Engineering Laboratory for Agri-product Quality Traceability, Beijing Technology and Business University (AQT-2022-YB3), the Key Technology Support Program of Taizhou City (TN202210) and the Key Research and Development Program of Zhenjiang City (SH2022006). The authors wish to thank the timely help given by Dr. Khalifa in improving the language of the article.

Data availability　Research data are not shared.

Conflict of interest　The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Funding information　National Key Research and Development Program of China (2016YFD0401104), Open Project Program of National Engineering Laboratory for Agri-product Quality Traceability, Beijing Technology and Business University (AQT-2022-YB3), the Key Technology Support Program of Taizhou City (TN202210) and the Key Research and Development Program of Zhenjiang City (SH2022006).

References

•  Abuzeina,  D. and  Al-Anzi,  F.S. (2018). Employing fisher discriminant analysis for Arabic text classification. Comput. Electr. Eng., 66, 474–486. doi: 10.1016/j.compeleceng.2017.11.002.
•  Ata,  Y.,  Baykal,  Y. and  Gokce.  M.C. (2019). Average channel capacity in anisotropic atmospheric non-Kolmogorov turbulent medium. Opt. Commun., 451, 129–135. doi: 10.1016/j.optcom.2019.06.055
•  Birks,  A.,  Green,  R.E., and  Mcintire,  P. (1991). Ultrasonic Testing in Non-Destructive Testing Handbook, vol. American Society for Non-Destructive Testing. doi: 10.3136/fstr.20.1033
•  Chaitanya,  G.M.S.K. and  Kumar,  A. (2021). Detection of subassembly bowing in a fast breeder reactor using glancing angle ultrasonic imaging. Ann. Nucl. Energy, 150, 107821. doi: 10.1016/J.ANUCENE.2020.107821
•  Cheng,  X.F.,  Zhang,  M., and  Adhikari,  B. (2014). Changes in quality attributes of strawberry purees processed by power ultrasound or thermal treatments, Food Sci. Technol. Res., 20, 1033–1041. doi: 10.3136/fstr.20.1033
•  Cho,  B.K., and  Irudayaraj,  J.M.K. (2003). Foreign object and internal disorder detection in food materials using noncontact ultrasound imaging. J. Food Sci., 68, 967–974. doi: 10.1111/j.1365-2621.2003.tb08272.x
•  Collado-Lara,  G.,  Heymans,  S.V.,  Rovituso,  M.,  Carlier,  B.,  Toumia,  Y.,  Verweij,  M,.  Paradossi,  G.,  Sterpin,  E.,  Vos,  H.J.,  D’hooge,  J.,  de Jong,  N.,  Van den Abeele,  K., and  Daeichin,  V. (2022). Spatiotemporal distribution of nanodroplet vaporization in a proton beam using real-time ultrasound imaging for range verification. Ultrasound Med. Biol., 48, 149–156. doi: 10.1016/j.ultrasmedbio.2021.09.009
•  Correia,  L.R.,  Mittal,  G.S., and  Basir,  O.A. (2007). Ultrasonic detection of bone fragment in mechanically deboned chicken breasts. Innovative Food Sci. Emerging Technol., 9, 109–115. doi: 10.1016/j.ifset.2007.06.004
•  Fortaleza,  L.G.S.,  Monteiro,  E.C.,  Barbosa,  C.R.H.,  Silva,  E.C., and  Gusmao,  L.A.P. (2018). Biomedical comparison of magnetometers for non-ferromagnetic met allic foreign body detection. J. Phys.: Conf. Ser., 1044, 012013. doi: 10.1016/J.MEATSCI.2020.108319
•  Geng,  Z.Q.,  Zhao,  S.S.,  Tao,  G.C., and  Han,  Y.M. (2017). Early warning modeling and analysis based on analytic hierarchy process integrated extreme learning machine (AHP-ELM): Application to food safety. Food Control, 78, 33–42. doi: 10.1016/j.foodcont.2017.02.045
•  Guo,  Q.,  Zhang,  L.,  Di,  Z.X.,  Ning,  C.P.,  Dong,  Z.W.,  Li,  Z.Y.,  Wang,  D.M.,  Liu,  C.,  Zhao,  M., and  Tian,  J.W. (2018). Assessing risk category of breast cancer by ultrasound imaging characteristics. Ultrasound Med. Biol., 44, 815–824. doi: 10.1016/j.ultrasmedbio.2017.12.001
•  Hæggström,  E. and  Luukkala,  M. (2001). Ultrasound detection and identification of foreign bodies in food products. Food Control, 12, 37–45. doi: 10.1016/S0956-7135(00)00007-4
•  Jiang,  Y.S.,  Zhang,  S.B.,  Zhang,  S.Y., and  Peng,  Y.X. (2021). Comparative study of high-intensity ultrasound and high-pressure homogenization on physicochemical properties of peanut protein-stabilized emulsions and emulsion gels. J. Food Process Eng., 44, 13710. doi:/10.1111/JFPE.13710
•  Joanna,  T.,  Sylwia,  K., and  Wojciech,  K. (2016). The Risk Analysis of Met allic Foreign Bodies in Food Products. J. Food Qual., 39, 398–407. doi: 10.1111/jfq.12193
•  Junior,  J.J.D.M.S. and  Backes,  A.R. (2016). ELM based signature for texture classification. Pattern Recogn., 51, 395–401. doi: 10.1016/j.patcog.2015.09.014
•  Kim,  W. (2019). X-ray Image Training Model for Enhanced Foreign Body Detection in Met als. J. Korea Conv. Soc., 10, 15–21. doi: 10.15207/JKCS.2019.10.10.015
•  Koizumi,  M.,  Ihara,  F.,  Yaginuma,  K.,  Kano,  H., and  Haishi,  T. (2010). Observation of the peach fruit moth, Carposina sasakii, larvae in young apple fruit by dedicated micro-magnetic resonance imaging. J. Insect Sci., 10, 1–10. doi: 10.1673/031.010.14105
•  Kumar,  V., and  Kumar,  C.P.S. (2020). Investigation of the influence of coloured illumination on surface texture features: A Machine vision approach. Meas., 152, 107297. doi: 10.1016/j.measurement.2019.107297
•  Kwak,  D.H.,  Son,  G.J.,  Park,  M.K., and  Kim,  Y.D. (2021) Rapid Foreign Object Detection System on Seaweed Using VNIR Hyperspectral Imaging. Sens., 21, 5279. doi: 10.3390/s21165279
•  Leemans,  V. and  Destain,  M.F. (2009). Ultrasonic internal defect detection in cheese. J. Food Eng., 90, 333–340. doi: 10.1016/j.jfoodeng.2008.06.042
•  Li,  C.Y.,  Cong,  R.M.,  Guo,  C.L.,  Li,  H.,  Zhang,  C.J.,  Zheng,  F., and  Zhao,  Y. (2020). A parallel down-up fusion network for salient object detection in optical remote sensing images. Neurocomputing, 415, 412–420. doi: 10.1016/j.neucom.2020.05.108
•  Li,  W.C.,  Frederick,  H., and  Palardy,  G. (2021). Multifunctional films for thermoplastic composite joints: Ultrasonic welding and damage detection under tension loading. Compos. Part A-Appl. S., 141, 106221. doi: 10.1016/J.COMPOSITESA.2020.106221
•  Li,  X.M.,  Han,  Z.W.,  Zhao,  T.Y.,  Zhang,  J.L., and  Xue,  D. (2021). Modeling for indoor temperature test based on time-delay and Elman neural network in air conditioning system. J. Build. Eng., 33, 101854. doi: 10.1016/J.JOBE.2020.101854
•  Liu,  Y.H.,  Wu,  J.Y.,  Chong,  C.J., and  Miao,  S. (2014). Ultrasound Assisted Osmotic Dehydration as Pretreatment for Hot-air Drying of Carrot, Food Sci. Technol. Res., 20, 31–41. doi: 10.3136/fstr.20.31
•  Nielsen,  M.S.,  Lauridsen,  T.,  Christensen,  L.B., and  Feidenhans’l,  R. (2013). X-ray dark-field imaging for detection of foreign bodies in food. Food Control, 30, 531–535. doi: 10.1016/j.foodcont.2012.08.007
•  Payel,  G.,  Sandeep,  S.R.,  Ankita,  N., and  Rama  C.P. (2016). Quality Evaluation of Food by Thermal Imaging. Int. J. of Proc. Post Harv. Technol., 7, 126–133. doi: 10.15740/HAS/IJPPHT/7.1/126-133
•  Ravikanth,  L.,  Singh,  C.B.,  Jayas,  D.S., and  White,  N.D.G. (2016). Performance evaluation of a model for the classification of contaminants from wheat using near-infrared hyperspectral imaging. Biosyst. Eng., 147, 248–258. doi: 10.1016/j.biosystemseng.2016.04.001
•  Rashmi,  S.,  Rekha,  A.,  Ashwani,  K.D., and  Anil,  A. (2021). Analytical Modelling of Hexagonal Shaped Capacitive Micromachined Ultrasonic Transducer. Int. J. Syst. Assur. Eng., 12, 252–262. doi: 10.1007/S13198-020-01046-Y
•  Ropelewska,  E. and  Szwejda,  G.J. (2018). A comparative analysis of the discrimination of pepper (Capsicum annuum L.) based on the cross-section and seed textures determined using image processing. J. Food Process Eng., 44, 13694. doi: 10.1111/JFPE.13694
•  Shi,  Y.,  Wang,  Y.Y.,  Hu,  X.T.,  Li,  Z.H.,  Huang,  X.W.,  Liang,  J.,  Zhang,  X.A.,  Zheng,  K.Y.,  Zou,  X.B., and  Shi,  J.Y. (2023). Nondestructive discrimination of analogous density foreign matter inside soy protein meat semi-finished products based on transmission hyperspectral imaging. Food Chem., 411, 135431. doi: 10.1016/j.foodchem.2023.135431
•  Srivastava,  G.K.,  Andrés-Iglesias,  C.,  Coco,  R.M.,  Fernandez-Bueno,  I.,  Medina,  J.,  García-Serna,  J.,  Dueñas,  A.,  Rull,  F., and  Pastor,  J.C. (2020). Chemical compounds causing severe acute toxicity in heavy liquids used for intraocular surgery. Regul. Toxicol. Pharmacol., 110, 104527. doi: 10.1016/j.yrtph.2019.104527
•  Suchanek,  M.,  Kordulska,  M.,  Olejniczak,  Z.,  Figiel,  H., and  Turek,  K. (2017). Application of low-field MRI for quality assessment of ‘Conference’ pears stored under controlled atmosphere conditions. Postharvest Biol. Technol., 124, 100–106. doi: 10.1016/j.postharvbio.2016.10.010
•  Song,  S.P. and  Ni,  Y.J. (2018). Ultrasound Imaging of Pipeline Crack Based on Composite Transducer Array. Chin. J. Mech. Eng., 31, 1–10. doi: 10.1186/s10033-018-0280-z
•  Stankovic,  M.,  Pantic,  I.,  De Luka,  S.R.,  Puskas,  N.,  Zaletel,  I.,  Milutinovic-Smiljanic,  S.,  Pantic,  S., and  Trbovich,  A.M. (2016). Quantification of structural changes in acute inflammation by fractal dimension, angular second moment and correlation. J. Microsc., 261, 277–284. doi: 10.1111/jmi.12330
•  Stone,  D.E. and  Quiroz,  L.H. (2016). Ultrasound Imaging of the Pelvic Floor. Obstet. Gynecol. Clin. North Am., 43, 141–153. doi: 10.1016/j.ogc.2015.10.007
•  Sun,  Z.B.,  Li,  J.K.,  Wu,  J.F.,  Zou,  X.B.  Ho,  C.T.,  Liang,  L.M.,  Yan,  X.J., and  Zhou,  X. (2021). Rapid qualitative and quantitative analysis of strong aroma base liquor based on SPME-MS combined with chemometrics. Food Sci. Hum. Wellness, 10, 362–369. doi: 10.1016/J.FSHW.2021.02.031
•  Tang,  J.X.,  Zhou,  H.,  Wang,  T.K.,  Jin,  Z.X.,  Wang,  Y.L., and  Wang,  X.Y. (2022). Cascaded foreign object detection in manufacturing processes using convolutional neural networks and synthetic data generation methodology. J. Intell. Mfg., 34, 2925–2941. doi: 10.1007/s10845-022-01976-3
•  Voss,  J.O.,  Doll,  C.,  Raguse,  J.D.,  Beck-Broichsitter,  B.,  Walter-Rittel,  T.,  Kahn,  J., and  Thieme,  N. (2021). Detectability of foreign body materials using X-ray, computed tomography and magnetic resonance imaging: A phantom study. Eur. J. Radiol., 135, 109505. doi: 10.1016/j.ejrad.2020.109505
•  Wang  Y.,  Wang,  X.F.,  Wang,  Z.P.  Li,  X.F.  Li,  C., and  Xu,  B.C. (2020). Effects of different freezing methods on the quality of conditioned beef steaks during storage. J. Food Process. Preserv., 44, e14496.
•  Wang,  Y.Y.,  Yan,  J.K.,  Rashid,  M.T.,  Ding,  Y.H.,  Chikari,  F.,  Huang,  S.F., and  Ma,  H.L. (2021). Dual-frequency sequential ultraso und thawing for improving the quality of quick-frozen small yellow croaker and its possible mechanisms. Innovative Food Sci. Emerging Technol., 68, 102614. doi: 10.1016/j.ifset.2021.102614
•  Wang,  Y.Y.,  Yan,  J.K.,  Ding,  Y.H., and  Ma,  H.L. (2021). Effects of ultrasound on the thawing of quick-frozen small yellow croaker (Larimichthys polyactis) based on TMT-labeled quantitative proteomic. Food Chem., 366, 130600. doi: 10.1016/J.FOODCHEM.2021.130600
•  Xu,  B.G.,  Yuan,  J.,  Wang,  L.,  Lu,  F.,  Wei,  B.X.,  Azam,  R.S.M.,  Ren,  X.F.,  Zhou,  C.S.,  Ma,  H.L., and  Bhandari,  B. (2020). Effect of multi-frequency power ultrasound (MFPU) treatment on enzyme hydrolysis of casein. Ultrason. Sonochem., 63, 104930. doi: 10.1016/j.ultsonch.2019.104930
•  Xie,  C.Q.,  Yang,  C., and  He,  Y. (2017). Hyperspectral imaging for classification of healthy and gray mold diseased tomato leaves with different infection severities. Comput. Electron. Agr., 135, 154–162. doi: 10.1016/j.compag.2016.12.015
•  Yelampalli,  P.K.R.,  Nayak,  J., and  Gaidhane,  V.H. (2018). Medical image rigid registration using a novel binary feature descriptor and modified affine transform. IET Image Process., 12, 337–344. doi: 10.1049/iet-ipr.2017.0526
•  Zhang,  X.L.,  Yue,  X.J.,  Ma,  B.,  Fu,  X.,  Ren,  H.L., and  Ma,  M.H. (2021). Ultrasonic pretreatment enhanced the glycation of ovotransferrin and improved its antibacterial activity. Food Chem., 346, 128905. doi: 10.1016/J.FOODCHEM.2020.128905