2017 Volume 23 Issue 4 Pages 545-549
The severity of internal browning in apple cultivars is often evaluated subjectively, making it potentially unreliable, and a method for automatic evaluation is necessary in order to process many samples efficiently. The objective of this study was to propose a model for estimating subjective browning severity ratings (SBSRs) in scanned images of sliced apples that mimics mean expert judgments. We assessed SBSRs made by three expert observers for images of sliced apples. The results indicated that the experts' evaluations of internal browning were qualitatively similar, but not quantitatively equivalent. The proposed model estimates the mean SBSRs of experts as a percentage of the browning region of the total flesh. The browning regions were qualified using CIELAB color difference from the standard color. The model estimations were consistent with increasing browning during longer storage periods.
Fuji, which is a cross between Ralls Janet and Red Delicious, is a very popular apple cultivar. While Fuji apples have a very long shelf life, they also have a tendency for internal browning (IB). The possibility of fruit with IB is influenced by pre- and post-harvest factors: It is increased in fruit that is harvested later, is more mature and is subjected to longer storage periods. The effects of storage environment on product damage have been reported previously (i.e., CO2 levels in controlled atmosphere (CA) storage (Clark and Burnmesiter, 1999) and exposure duration (Voltz et al., 1998; Argenta et al., 2000)). Many researchers have performed studies to detect, predict or control the development of IB, and have used their own processes or developed their own methods to evaluate the degree of IB. For example, the number of fruits with browning in a given sample is often assessed to investigate group differences, such as of CO2 levels in controlled atmospheres and temperatures (Watkins, 2010) or 1-MCP treatment (Moran and McManus, 2005). A non-destructive method, such as using near infrared spectroscopy transmittance to identify IB (Clark et al., 2003), to inspect for IB is necessary, but a sufficiently accurate method has not yet been developed.
Some studies have evaluated the degree of IB using the browning index, which is based on scores of browning severity of individual fruit. Various procedures have been used to score browning: Buera et al., (1986) proposed a browning index based on the x-value of CIE 1931 for common foods containing sugar; however, this was developed for the Maillard reaction and not for internal browning. Hatoum et al. (2014, 2016) used a browning index defined as the mean square value of browning rate, which was scored using their own calibration cards, for flesh regions in slice photographs. However, they did not describe how they defined their calibration cards. The scores were typically determined by each researcher visually (i.e., 1 = healthy and 5 = severely bruised; de Castro et al., 2008; Vanoli et al., 2011, 2014; Lee et al., 2012), potentially rendering the scores unreliable, even if measuring the same fruit, due to subjective differences in ratings. However, no research has explored the reliability of such subjective evaluation of IB.
In addition to the need for greater objectivity, a method for automatic evaluation of browning severity is necessary in order to process many samples efficiently. In this study, we measured the subjective evaluation of IB by three expert observers and propose a model called the subjective browning severity rating (SBSR) to evaluate the IB of Fuji apples.
Materials Fuji apples (Malus domestica Borkh cv. Fuji) were used in this experiment. In order to collect samples with various degrees of IB to present to observers, fruits sorted by a sorting machine as having browning or not were included. All fruit was harvested in Aomori, Japan in November 2013 and stored at 0°C in controlled atmosphere conditions until July 2014. Before being placed into controlled atmosphere conditions, the fruit was stored in regular atmosphere at 0°C for 3 weeks after harvest to avoid overly severe CO2 injury. We sampled twenty apples in total: Six of these were stored in a controlled atmosphere with low CO2 levels (CO2 1.5%, O2 2.2%) and identified as having browning flesh using near infrared spectroscopy transmittance on the fruit-sorting machine (QSCOPE-F64HM-F; Mitsui Kinzoku Instrument Technology Corp., Komaki, Aichi, Japan). The sorting machine was a transmission-type model that allowed non-destructive measurements. However, in some cases the machine measurements produced errors, for example, when discolored portions are distributed spatially, when fruit was frozen or thawed, or with small fruits with large amounts of liquid. Another six samples were stored at low CO2 levels without browning, five were stored at high CO2 levels (CO2 5.0%, O2 2.2%) with browning, and three were stored at high CO2 levels without browning. To obtain sample images of IB levels, the fruits were brought out of storage and scanned at room temperature, 20 – 25°C. Each fruit was sliced at the equator at a thickness of 1 cm, and the slice was captured by a flatbed image scanner (GT-X980, Seiko Epson Corp., Suwa, Nagano, Japan). The flatbed scanner is readily available and allows the easy acquisition of highly stable samples because its shooting environment (i.e., the configuration of light source, camera and scanning surface) is fixed within the same product. The cover of the scanner (the transparency unit) was removed to accommodate the fruit; thus, in order to eliminate external lighting, the fruit was covered by a black Bakelite board during scanning. The scanning resolution was 1200 dpi and the color depth was 16 bits per R/G/B component.
Measurement of subjective browning severity rating To investigate the subjective evaluation of IB, we measured the SBSR. In this study, the SBSR was defined as the percentage of an area occupied by browning flesh. The SBSR represents the degree of IB as a continuous quantity; therefore, it is useful when a detailed investigation of each individual fruit is required. To display the same fruit under the same conditions for several expert observers, we measured the ratings for slice image patches. Slice patches were cropped to 86.7 × 86.7 mm regions from scanned images, then resized to one-tenth resolution because the original image resolution was too high to display (Fig. 1). Internal browning is an important issue in agricultural and sensory evaluation food studies. Therefore, three expert observers participated; one was a doctor of agriculture who was studying diseases of apples, and the others had been participating for more than a decade in sensory evaluation research of food. The observers' task was to evaluate the area (%) of browning flesh for each slice; for example, if half the area of a slice was occupied by browning, they should input 50%.
Scanned slice images of Fuji apples. The slice images were sorted by mean SBSRs provided by expert subjects. The images go from low SBSRs (left side) to high SBSRs (right side).
Since the aim of the experiment was to evaluate the IB of slice images, differences in the controlled atmosphere conditions of fruit with/without browning were not analyzed. Figure 2 shows the SBSRs for each individual slice image for all three expert observers. The interclass correlation coefficient was high (ICC(3,1) = 0.85), indicating that the SBSRs had high intra-subject reliability. This suggests that the three experts' ratings for browning were qualitatively similar. A two-way ANOVA without replication was conducted to test the effect of expert and fruit. Each main effect was significant for expert [F(2,38)=15.04, p < 0.01] and for fruit [F(19,38)=17.62, p < 0.01]. Using paired t-tests with Bonferroni's correction for post hoc analysis, we revealed that expert C rated higher than experts A and B. The results revealed that the experts' evaluations of the areas with IB were qualitatively similar (high ICC), but not quantitatively equivalent (expert C rated higher than the others). Therefore, we used the mean SBSR of the three experts for each individual fruit in model estimations.
Evaluated subjective browning severity ratings for slice images by three expert subjects. The SBSRs for each slice image are ordered in number in Fig. 1.
Model We propose a model for estimating the SBSR from scanned images of sliced apples (Fig. 3). In this model, pixels with a substantial color difference from the standard color are defined as browning, and the percentage of browning in the flesh area is determined. It should be noted that IB was examined only for the flesh of the apple slices, not for the core, as we did not target the core flesh for this experiment. An in-house developed MATLAB code was used to estimate SBSRs from the slice images.
Flowchart of the model for estimating SBSRs from scanned slice images.
First, in order to crop the flesh region of images, the slices were approximated into circles using the circle Hough transform (CHT) technique: The slice images were resized to one-sixteenth resolution to hasten the process, then converted to binary images for CHT. Beforehand, we visually measured eccentricities of the vascular bundles for 154 fruit slices, and observed that the vascular bundles were located at between 35% and 65% of the radius (mean: 51.0%, standard deviation: 7.4%), hence the area within 65% of the radius was treated as the core in this model. The area near the edge of each slice was excluded because it was thin, causing the scanned images to be dark. Therefore, rings with radii of 67% to 95% were identified within the circular flesh slices.
The scanned images were converted to CIELAB color space from RGB values because the SBSR is based on a subjective visual judgment. CIELAB is designed to approximate human color perception, in which L* indicates lightness, a* is the red-green coordinate and b* is the yellow-blue coordinate. In addition, Lee et al., (2012) suggested that L* and hue angle (tan−1(a/b)) changes in the fruit flesh are correlated with the development of flesh browning.
We defined a standard color ([L*s, a*s, b*s]) and color difference threshold (κ). The color differences (ΔE*) from the standard color were calculated for each flesh pixel.
 |
where L*i, a*i and b*i indicate i-th pixel's color. Then, pixels with a color difference over the threshold (κ) are treated as browning
 |
 |
Model results. (a) The image on the right indicates the model estimation for the slice image to the left. The red pixels indicate browning as proposed by the model. The cyan circle indicates a radius of 0.67, and outside of that radius is the flesh region targeted for this study. The SBSR is the percentage of red (bruised) area to total flesh in the target area. (b) Comparison for each apple of the SBSRs estimated by the model, which fit to the mean SBSRs among experts (open diamonds), and those measured by each expert.
We calculated the model parameters to fit the mean SBSRs among experts. Four model parameters, three components (L*s, a*s, b*s) of standard color and the color difference threshold (κ), were estimated by applying the Monte Carlo method ([L*s, a*s, b*s] = [75.6, −15.3, 27.4], κ = 15.1) for least squares fitting to SBSRs for the 20 samples used in the experiment. Figure 4(b) shows the SBSRs estimated by the model and measured by expert comparison for each apple. The correlation coefficient between measured mean SBSRs and estimated SBSRs was 0.98, demonstrating that the proposed model provides a reasonable evaluation of SBSRs for scanned images of sliced apples. Even when the model parameters were adjusted to the mean SBSRs, the correlation coefficients between the SBSR of each expert and the model estimation were sufficiently high: 0.98 for expert A, 0.90 for expert B and 0.92 for expert C.
Model applications We applied the model to estimate SBSRs of Fuji apples with different storage periods. Typically, development of IB occurs with long durations of storage. All fruits were harvested in Aomori, Japan in November 2014, and stored at 0°C for 4, 5, 6 or 8 months. Forty apples for each storage period were cut and scanned as with the previous experiment, and the SBSRs were calculated using the proposed model. Figure 5 shows the estimated SBSRs for each storage period. An ANOVA revealed a significant effect of storage period (F(3,36)=13.01, p < 0.05). The samples stored for 8 months were estimated to have higher SBSRs than those of the other storage periods. These model estimations were consistent with the tendency of IB to increase with longer storage periods.
Model application results. The estimated SBSRs for different storage periods are shown; the dashed lines indicate first and third quartiles, the error bars indicate lowest and highest datum still within a 1.5 interquartile range. *SBSR outliers.
We measured the SBSRs of three expert observers, and the results suggested that SBSRs can differ due to individual differences despite similar tendencies. To further understand the internal browning of Fuji apple cultivars, we proposed a model to estimate SBSRs from images of sliced apples. The measured SBSRs were well represented by our model, which was based on percentages of pixels with a substantial color difference in flesh area. When we applied the model to apples with different storage periods, the estimations were consistent with the tendency of IB to increase with longer storage periods. These results suggest that the model is reasonably equivalent to estimated SBSRs by expert observers. Since the model provides stable evaluation of IB even if experts are not present, it should be useful when analyzing large quantities of samples simultaneously.
