Original papers
Food Texture Quantification Using a Magnetic Food Texture Sensor and Dynamic Time Warping
2018 年 24 巻 2 号 p. 257-263
詳細
2018 年 24 巻 2 号 p. 257-263
Food texture is an important characteristic related to food preferences. Food texture instruments are used to determine the physical profiles of foods; however, they are not sufficient for detailed evaluation of food texture. In this study, a novel method for the quantitative evaluation of food texture is proposed. The proposed method records time-series data of force and vibration in fractures for different foods that have been estimated to have a similar food texture. Then, the dynamic time warping barycenter averaging algorithm is used to determine the standard data based on the recorded time-series data. Finally, the dynamic time warping algorithm is used to calculate the similarity between the measured data and the standard data, which is regarded as the quantitative food texture. The effectiveness of the method was confirmed with experimental results.
The intake of food is an essential behavior that is necessary for the maintenance of human life and health. Humans perceive mainly three stimuli when eating: taste, smell, and food texture. Taste and smell are classified as chemical characteristics, while food texture is classified as a physical characteristic. Thus, food texture is an important factor that has a great effect on the palatability of foods. Furthermore, the Japanese have numerous food texture expressions compared to other languages (Nishinari et al., 2008), indicating that the Japanese place particular importance on food texture. Therefore, food manufacturers need to quantitatively evaluate food texture to develop foods with high consumer appeal.
Sensory evaluation is the typical method used for food texture evaluation; however, this method has some limitations. First, it tends to be influenced by the age, gender, and living environment of the human subjects. Additionally, it is difficult to quantitatively evaluate human subject data because the results are subjective and qualitative. Moreover, sensory evaluation requires large numbers of human subjects to make statistically significant inferences, and thus is resource and time intensive.
On the other hand, we can quantitatively evaluate food texture using a food texture instrument. Food texture instruments measure the time-series load by pressing a food sample with a plunger in the vertical direction. Szczesniak analyzed the time-series load and proposed a texture profile analysis that calculates various physical properties, such as hardness, cohesiveness, and adhesiveness. These properties are calculated from the time-series load as measured by a food texture instrument (Szczesniak, 1963). Texture profile analysis became the standard method of food texture evaluation and remains in use today. However, food development technologies in recent years have resulted in more complicated food textures, and texture profile analysis is no longer considered comprehensive. Evaluation of physical properties alone is not sufficient for such complex textures. Therefore, new evaluation methods, such as the acoustic signal method, have been devised (Taniwaki et al., 2010; Taniwaki and Kohyama, 2012).
The purpose of this study is to establish a new method that qualitatively evaluates human perceptions of food texture. Humans use a wide variety of expressions to describe food texture. This study develops an evaluation system based on food texture expressions rather than physical properties. The system evaluates food textures according to the similarity of time-series data measured by a food texture sensor. A similarity is calculated from the standard data and the measured data of time-series data, and it is regarded as the evaluation value of food texture. To calculate similarities, we use the dynamic time warping algorithm (DTW) and the DTW barycenter averaging algorithm (DBA).
Materials A magnetic food texture sensor was used in this study (Nishikubo et al., 2016; Figure 1). This sensor comprises a contactor, an elastomer, a base, and a sensor board. The material of the elastomer is urethane, and that of the contactor is ABS resin. The surface area of the contactor is a circular plane with a radius of 5.0 mm, and the elastomer is arranged with a width of 5.0 mm around it. The contactor contains a permanent magnet and acts as a tooth. The elastomer acts as the periodontal membrane, and the base acts as the alveolar bone. Since the elastomer is a material that expands and contracts, the contactor moves up and down according to the food's force.
Photograph of the texture sensor.
Figure 2 shows a system diagram of the food texture sensor. A personal computer (PC) sends a serial communication to a z-axis motorized stage (SGSP26-100, SIGMAKOKI Co., Ltd., Saitama, Japan), and moves the food texture sensor to compress the food. Based on the change in magnetic field strength by the displacement of the contactor, the food texture sensor measures the output voltage of the giant magnet resistance (GMR) effect element, and the induced voltage is produced by an inductor. These voltages are amplified and undergo analog to digital conversion in the PC.
System diagram of the food texture sensor.
The GMR element in the food texture sensor outputs static voltages based on the magnetic-field strength in the sensor board, and it tends to be influenced by manufacturing errors or the dispersion of each element. Therefore, it is necessary to calibrate the output data of the GMR elements. In this sensor, we measured a load using a force sensor while measuring output voltages of the GMR elements, and performed the calibration using a multiple regression analysis.
Methods As mentioned in the Introduction section, we evaluated the food textures by calculating the similarity of data from the food texture sensor. Figure 3 shows a flow chart of the food texture evaluation in this study. First, the standard data of food textures are determined by sensory evaluation and the measured result of the food texture sensor. Next, the similarity between the standard data and the measured data is calculated, and this is regarded as the evaluation value of food texture. Generally, the Euclidean distance is used as the similarity calculation method. However, the time-series data by the food texture sensor were corrupted by divergences in the time direction and signal strength direction. Therefore, it was not possible to calculate the similarity with the Euclidean distance because that data must be included in the same dimension. To solve this problem, we used the DTW to evaluate the data. The DTW is an algorithm used for measuring the similarity between two time-series data sets, which may vary in speed. The DTW is also used for image and sound recognition (Lindasalwa et al., 2010; Giorgio et al., 2004). The merit of using the DTW is its use for non-linear data and its ability to set expansion and contraction limits, and start and end points.
Flow of the proposed evaluation method
Consider the following two time-series data sets: Eqs. 1 and 2.
 |
 |
 |
 |
 |
 |
To perform a quantitative evaluation using the DTW, this method needs the standard data of each food texture, as shown in Figure 3. In this study, we select a food that seems to have a specific texture to create data based on sensory evaluation, and define the measured data of the food as the standard data of texture. However, it is difficult to determine the standard data from just one food. Hence, we select multiple foods for sensory evaluation, and determine the standard data from the time-series data of those foods. In this way, we first determine the target textures, and then select various foods having those target textures. To average several measured data, we used the DBA, which is an averaging algorithm that uses the DTW calculation (Petitjean, 2011). The DBA determines time-series data that has the divergence in the time axis as DTW can. The DBA involves the five following steps.
Experiment To confirm the effectiveness of the system, a fundamental experiment was conducted. First, we decided upon two target food textures: “mochimochi” and “mocchiri.” Although both “mocchiri” and “mochimochi” include the food textures of chewy and sticky, they are slightly different. However, most Japanese can distinguish the difference. In this paper, “mochimochi” is defined as the texture in which we continuously feel good elasticity and stickiness in our mouth, and “mocchiri” is defined as the texture in which we feel elastic softness at the moment it is placed in the mouth. These textures are currently in high demand by consumers according to a questionnaire survey performed by the BMFT Company (Ohashi et al., 2015). In addition, since the DTW cannot deal with irregularly shaped data, we chose these textures because they are less crushed and can be effectively applied to DTW. We selected nine commercially available foods that have the texture of “mochimochi” or “mocchiri,” or are very similar to these textures (Hayakawa et al., 2012). The selected foods are shown below.
Measurement methodology using the proposed sensor First, we performed the measurement experiment using the food texture sensor. The mold height required for insertion into the food sensor was 20 mm, and each sample was cut to this height and placed on the stage of the machine. The bracken-starch dumpling and jelly could not be cut; thus, we used their original height (the bracken-starch dumpling was 18 mm and the jelly was 25 mm). We suggest that their heights did not significantly affect the measured force because their elasticity is lower than the other foods. Then, to push round-trip, we used circular interpolation of the x-axis and the z-axis of the device, which was similar to simple harmonic motion. The initial velocity was set at 5.0 mm/s and the pushing depth was set at 15 mm. The push frequency was set to make a round-trip twice. Although the velocity is slower than the speed of human biting, the DTW relatively evaluates the difference between the time-series data by measuring at the same velocity of 5.0 mm/s. Each food was measured 10 times. Figure 4 shows the time-series force and vibration derived from the magnetic food texture sensor for each food.
Force and vibration by the food texture sensor.
Sensory evaluation and comparison We performed a sensory evaluation of the nine selected foods concomitant with the measurement experiment. In the sensory test, 10 panelists (9 male and 1 female students in their twenties) were selected to evaluate whether each food was either “mochimochi” or “mocchiri” on a scale of 1 (I can not feel that texture in the food) to 5 (I can feel the texture prominently in the food). We gave samples of each food to panelists in random order to counteract the effect of the order of proffer. We calculated the average of sensory values for “mochimochi” and “mocchiri”, and divided the foods into two groups depending upon whether the stated “mochimochi” value was larger than the “mocchiri” value. Table 1 shows the difference in values for each food sample and whether the food was considered “mochimochi” or “mocchiri.”
| Food | “mochimochi” | “mocchiri” | “mochimochi”–“mocchiri” | Group |
|---|---|---|---|---|
| Bracken-starch dumpling | 4.3 | 3.5 | 0.8 | Group “Mochimochi” |
| Rice cake with kinako | 4.6 | 4.0 | 0.6 | |
| Dumpling | 4.5 | 4.0 | 0.5 | |
| Donut | 3.7 | 3.3 | 0.4 | |
| Baguette | 2.2 | 2.2 | 0.0 | Even |
| White bread | 2.4 | 2.7 | −0.3 | Group “Mocchiri” |
| Marshmallow | 3.1 | 3.4 | −0.3 | |
| Jelly | 1.1 | 1.6 | −0.5 | |
| Boiled fish-paste | 1.6 | 2.7 | −1.1 |
Determining the standard data for each food and an evaluation of DTW distance Based on the two groups according to the sensory evaluation results, standard time-series data were constructed using averaging by the DBA. We regarded these standard data as time-series data of “mochimochi” and “mocchiri.” Figure 5 shows these two time-series data. Next, the DTW distance between the standard data of each texture and the measured data of each food was calculated. At this time, DTW distances regarding vibrations were not calculated because we suggest that there was no difference in the food texture due to vibrations, since the induced voltages by the inductor were not generated. Figure 6 is a scatter diagram of the DTW distances regarding each texture.
Averaging data “Mochimochi” and “Mocchiri” using DBA.
DTW evaluation value for “Mochimochi” and “Mocchiri.”
Figure 4 shows the force from foods to the contactor calculated by the output voltage of GMR elements, and the induced voltages caused by the inductor. Induced voltages were not generated by all the foods. We suggest that vibrations were not generated because “mochimochi” and “mocchiri” are textures that do not require crushing by the teeth. The GMR's graph showed that there was no difference between the first peak value and the second peak value. This tendency is often seen in foods that have large values of springiness and cohesiveness. Nearly all the foods we selected were considered to have a large amount of springiness. The results of this measurement show that the forces from the rice cake with kinako, the dumpling, the donut, and the boiled fish-paste, as measured by the GMR elements, are higher than the other foods. Conversely, the forces from the jelly and the marshmallow are lower. In comparison with the sensory evaluation results shown in Table 1, the force in the “mochimochi” group tended to be larger than that in the “mocchiri” group. This tendency can be seen in Figure 5, where the “mochimochi” graph illustrates a larger force than the “mocchiri” graph.
Figure 6 shows that the differences in the DTW distance can be detected for each food. The dumpling had the lowest DTW distance value of “mochimochi,” while the marshmallow had the lowest DTW distance value of “mocchiri.” To facilitate data comparison with the sensory evaluation, we divided the data into two groups, i.e., the same groups that were used for the sensory evaluation. We then examined the difference in DTW distances between “mochimochi” and “mocchiri,” as shown in Figure 6. The grouping of data by DTW evaluation almost coincided with the grouping by sensory evaluation, with the exception of the bracken-starch dumpling and the baguette. The DTW distance of the bracken-starch dumpling was categorized as belonging to the “mocchiri” group. The group differed from the result of the sensory evaluation. This difference was because the standard data of “mocchiri” is largely reflected from the jelly's force data. Hence, foods with smaller force values tended to be grouped into “mocchiri”. This result indicates that the method using the DTW needs to exclude foods with radically different tendencies, such as jelly, from determination of the standard data. The method of the determination should be studied further in future works. Although the sensory evaluation of the baguette was “Even”, the DTW distance was grouped into “mochimochi”. This is attributed to the low accuracy of the evaluation value because of the small number of panelists in the sensory evaluation.
We can therefore conclude that the food textures calculated using the DTW distance and that calculated using the magnetic food texture sensor demonstrate the same trend as with the human perception of food texture. As mentioned in the Introduction section, since Szczesniak's texture profile uses only a part of the time-series data, it is difficult to distinguish and evaluate textures, such as “mochimochi” and “mocchiri,” that have similar parameters. The method proposed herein uses the entire time-series data, and evaluates differences between such similar textures. The similarities are calculated directly as numerical values by the DTW. Hence, the proposed method is considered to be effective as a quantitative evaluation method of food texture.
A new evaluation method was proposed that evaluates food texture quantitatively using the DTW and DBA. We performed measurement and sensory tests to investigate whether this method can measure sensory data that are similar to the sensory experiences reported by humans. From the experimental results, this method is considered to be effective as a novel quantitative evaluation method of food textures. Future research will apply this method to other textures to ensure a wider scope of applicability. To determine appropriate standard data, the determination method of standard data should be discussed further. In addition, future research will be directed at examining this method in cases where vibration may be a factor in food texture. This aspect of the research will utilize the advantages of the magnetic food texture sensor.
The practical application of this evaluation method may foster advances in food technology. In the design phase of foods, this method can be used to match texture requirements by quantitatively establishing the degree of difference or similarity in the texture of the product compared to a target food texture. Also, in the food development industry, there is a great need for quantitative evaluation during inspections, and this evaluation method can be utilized throughout the food manufacturing process.
Acknowledgements This work was supported by JSPS KAKENHI Grant Number JP16K00813.
