In Search of Optimum Fresh-Cut Raw Material: Using Computer Vision Systems as a Sensory Screening Tool for Browning-Resistant Romaine Lettuce Accessions

Abstract

The popularity of ready-to-eat (RTE) salads has prompted novel technology to prolong the shelf life of their ingredients. Fresh-cut romaine lettuce is widely used in RTE salads; however, its tendency to quickly discolor continues to be a challenge for the industry. Selecting the ideal lettuce accessions for use in RTE salads is essential to ensure maximum shelf life, and it is critical to have a practical way to assess and compare the quality of multiple lettuce accessions that are being considered for use in fresh-cut applications. Thus, in this work we aimed to determine whether a computer vision system (CVS) composed of image acquisition, processing, and analysis could be effective to detect visual quality differences among 16 accessions of fresh-cut romaine lettuce during postharvest storage. The CVS involved a post-capturing color correction, effective image segmentation, and calculation of a browning index, which was tested as a predictor of quality and shelf life of fresh-cut romaine lettuce. The results demonstrated that machine vision software can be implemented to replace or supplement the scoring of a trained panel and instrumental quality measurements. Overall visual quality, a key sensory parameter that determines food preferences and consumer behavior, was highly correlated with the browning index, with a Pearson correlation coefficient of −0.85. Other important sensory decision parameters were also strongly or moderately correlated with the browning index, with Pearson correlation coefficients of −0.84 for freshness, 0.79 for off odor, and 0.57 for browning. The ranking of the accessions according to quality acceptability from the sensory evaluation produced a similar pattern to those obtained with the CVS. This study revealed that multiple lettuce accessions can be effectively benchmarked for their performance as fresh-cut sources via a CVS-based method. Future opportunities and challenges in using machine vision image processing to predict consumer preferences for RTE salad greens is also discussed.

1. Introduction

Ready-to-eat (RTE) salads are minimally processed and highly perishable products that are more vulnerable to quality degradation compared to their whole vegetable counterparts [1]. Quality (e.g., flavor, color, texture) degradation in fresh-cut produce can be attributed to multiple factors, including biochemical reactions, microbial spoilage, and physiological changes [2,3,4,5,6,7,8].

In the United States, sales of RTE salads became prominent in the produce sections of retail outlets over two decades ago [9,10]. The popularity of these products has coincided with increased demand for romaine lettuce [11]. The growing consumer interest and craving for romaine lettuce has presented challenges for fresh-cut processors, given the tendency of romaine lettuce to rapidly develop brown tissue on cut surfaces [12,13].

Postharvest technology has been developed to slow down the perishability of RTE leafy greens. This has included the identification of adequate polymer films and storage temperatures to ensure optimum gas composition in the passive/active modified atmosphere produced in the headspace of the packages [14,15,16,17,18,19], novel processing technology to cut and drain the product [20], and sanitizing systems that can prolong quality [21,22,23]. One of the latest approaches that aims to maximize postharvest quality across the supply chain is the development of improved breeding lines [3,24]. While traditional vegetable breeding programs have focused on improved disease tolerance and harvest yields, the growing sophistication of consumers and the evolution of RTE products have prompted more attention to breeding programs that target postharvest shelf life and RTE quality [25,26,27]. For example, diverse accessions of romaine lettuce with improved shelf life in modified atmosphere packaging (MAP) are being developed [19,28,29].

Postharvest quality of romaine lettuce is highly dependent on the degree of browning. Brown tissue is caused by the oxidation of polyphenols, accelerated by enzymatic reactions that involve polyphenol oxidase (PPO), phenylalanine ammonia lyase (PAL), and peroxidase (POD) [3,8]. Browning of fresh produce is critical in driving consumer behavior. It is widely accepted that consumers make purchasing decisions on RTE products based on the overall visual quality and extent of browned tissue [30], while likability of the product may be based on secondary factors, such as freshness and green color [31,32].

Postharvest performance of lettuce is affected by multiple factors, including its genetic constitution inherited from the parents and the environment during production, processing, and storage. Because of this extensive variability within and across production seasons, it is necessary to determine effective ways to assess the quality of the final product [19,29,33]. The need for a rapid and reliable way to predict consumers perception for quality and their preferences has become a priority for researchers and sensory professionals. The goal is to reduce time and its associated costs in implementing traditional methods that rely on instrumental quality measurements and trained panelists or large consumer panel assessments [23,34,35]. This goal can be achieved by substituting human involvement in sensory evaluations with approaches based on computer vision systems (CVS). Advancements in digital technology and CVS, which includes highly accurate segmentation of images [36,37,38,39,40,41], provide a promising outlook for rapid and accurate assessment of postharvest quality.

A wide variety of machine learning-based methods for image processing have recently emerged that can potentially be used for assessing the quality of fresh produce. These approaches have started to be tested on leafy greens, including lettuce. For example, a CVS fed with a machine learning method (support vector machine) was recently used with 70% effectiveness in predicting when spinach would wilt using CIELAB color space to recognize changing color during postharvest storage [42].

The objective of this study was to determine whether digital image analyses can be effectively used to screen raw romaine lettuce cultivated at different growing seasons. To do so, the precision of the researched method was tested on a set of commercial lettuce cultivars and newly developed breeding lines.

2. Materials and Methods

2.1. Romaine Lettuce Processing and Packaging

Romaine lettuce was grown at the U.S. Department of Agriculture, Agricultural Research Service, Sam Farr United States Crop Improvement and Protection Research Center, in Salinas, CA, USA, using the standard agronomical practices for the area. A group of 15 accessions was evaluated (Figure 1) and consisted of 12 commercial cultivars (Braveheart, Clemente, Green Forest, Green Towers, Hearts Delight, King Henry, Lobjoits, Parris Island Cos, Siskiyou, Sun Valley, Tall Guzmaine, and Triple Threat), 2 breeding lines (RH11-1506 and SM13-R2), and a single plant introduction (PI 491224). After harvesting, the lettuce was cooled to 5 °C with forced air cooling and shipped to the USDA Food Quality Laboratory (FQL) at the Beltsville Agricultural Research Center (BARC) in Beltsville, MD, USA, in a commercial refrigerated truck (2–4 °C) within 3 d of harvest. Upon arrival at the FQL, the lettuce was transferred to a 5 °C cold room for 1 d before processing, packaging, and quality assessment.

Outer leaves were manually removed from each lettuce head, and the stems and tops were trimmed off. The trimmed lettuce was cut into 2.5 × 2.5 cm squares using a pilot-scale cutter (Nichimo Seven Chefs ECD-302, Tokyo, Japan), washed for 30 s in a 100 mg L⁻¹ NaClOH adjusted to pH 6.5 with HCl (1 mol L⁻¹), and centrifugally dried for 2 min at 666 rev min⁻¹ using a pilot-scale dryer (Model T-304, Meyer Machine Co., San Antonio, TX, USA). Lettuce (128 g) was packaged in 22.9 × 15.3 cm polypropylene bags (Packaging Concept Inc., Salinas, CA, USA) with a film oxygen transmission rate at 5 °C of 16.6 pmol O₂ s⁻¹ m⁻² Pa⁻¹ [17]. Packaged lettuce (4 replicates.) was stored at 5 °C for up to 14 d for quality assessment.

2.2. Sensory Evaluation

Lettuce sensory evaluation was performed by six trained panelists recruited from BARC. The panel members were 50% female and 50% male, ranging from 25 to 55 years of age. Before sample evaluation, the panelists were trained to visually assess differences in freshness, browning, off-odor (olfactory evaluation), and the overall visual quality (OVQ) using fresh-cut lettuce samples with diverse degrees of the targeted quality parameters [23]. All quality parameters were rated on a 0–100 scale. Freshness was rated based on the appearance of the sample, with 100 being the freshest. Samples with the lowest rating referred to cuts that were limp, wilted, discolored, damaged, and/or decayed [23]. Browning was rated based on the appearance of the cut edge, excluding decay or decomposition (soaked dark tissue), with a score of 0 linked to the absence of browning tissue. During the training period, the sensory panel facilitator worked with panelists to develop a pictorial intensity scale for browning (Figure 2) that was used as a reference during sample evaluation. Off-odor ratings were determined by sniffing the cuts and detecting the presence of unusual aroma, with a score range in which 0 was equivalent to no off odor. OVQ was rated based on the general appeal of the lettuce sample, combining all potential aspects that could form the sensory quality of the product, with 100 indicating an outstanding or maximum overall visual quality.

Sensory evaluations were performed inside a temperature-controlled room (22 °C) with incandescent lighting. The lettuce samples were rated for freshness, browning, and overall visual quality 3, 7, and 14 days after processing. Off odor was only rated 14 days after processing. Any further comments from the trained panelists about the samples were also recorded. The panelists evaluated 15 samples in a randomized order during each of the sessions. The lettuce samples were removed from the original packaging and placed on white trays labeled with random three-digit numbers. In between samples, the panelists sniffed coffee beans to recalibrate their sense of smell [43,44]. One sample included the entire contents of a lettuce package (128 g), and four replicate samples of each lettuce accession were evaluated at each time point by each panelist. In total, 12 separate evaluation sessions were organized to conduct the sensorial quality assessments of all the lettuce samples in this study.

2.3. Color and Browning Assessment

Lettuce sample color and browning severity was quantified using a computer vision system (Figure 3A) including a computer, camera, and controlled lighting inside an enclosed box to block ambient light, and subsequent image post-processing and analysis [8,29,45]. The controlled lighting consisted of two sets of light-emitting diodes (LEDs) (5600 K daylight balanced, high color-rendering index) with diffusers (Amazon Basics Portable Photo Studio, Amazon, Seattle, WA, USA). Each lettuce sample (4 replicates) was photographed on a matte white background by spreading out the contents of a package (128 g) across a 40.5 × 21.5 cm area. As described by Teng et al. [8], image acquisition software (Nikon Camera Capture 2.0, Nikon Inc., Melville, NY, USA) was used with a laptop computer (Dell Precision 7720, Dell Inc., Round Rock, TX, USA) to control the camera (Nikon D 800, 60 mm lens, Nikon Inc., Melville, NY, USA) with settings (F-20, 1/30 s, ISO 640, high-quality raw file format (.nef)) that were determined to be optimal using a 24 color reference card (X-rite ColorChecker^® Passport, X-rite Inc., Grand Rapids, MI, USA).

Image post-processing started with a color correction using the previously described color card with the ColorChecker Camera Calibration software (v1.1.2, X-rite Inc., Grand Rapids, MI, USA) and Adobe Lightroom (version 6.3, Adobe Systems Inc., San Jose, CA, USA), followed by exporting photographs in a high-quality (uncompressed) .tiff format [8,29]. Images were segmented and analyzed using Image Pro Premier software (v.9.3b, Media Cybernetics, Inc., Rockville, MD, USA) with the smart segmentation tool. This machine learning software tool was used to develop a new method to segment lettuce images by classifying pixels into categories based on defined reference objects for each category. The categories in this method included background areas, regular pieces (dark green leaves, ribs, and inner leaves), and browned pieces (light, medium, dark) (Figure 3B). The segmentation method was created using images from preliminary trials, and the same method was then applied to all the sample photographs in this study. When analyzing a sample photograph, each pixel in the image was compared to the values of the reference object. Then, a multi-dimensional weighted Euclidian distance was calculated between the image pixel value and the reference values, allowing the pixel to be assigned to the class of the closest reference object using the software program.

The number of pixels in each category and the average RGB color of each category was calculated with the Image Pro Premier software (v.9.3b) and converted to L*a*b* (CIELAB 1976) with MATLAB (“rgb2lab” function, Image Processing Toolbox, v.R2017b, MathWorks, Natick, MA, USA). Following that, the values of hue angle (h°) and chroma (C) were calculated from a* and b* with standard equations hue angle = tan⁻¹ [(b*) (a*)⁻¹] and chroma = [(a*)² + (b*)²]^0.5 [46,47,48].

The overall weighted average color (L*, a*, b*, or h°) of each lettuce sample was calculated by combining the average colors from each category based on their relative amounts:

$$\mathrm{W}\mathrm{e}\mathrm{i}\mathrm{g}\mathrm{h}\mathrm{t}\mathrm{e}\mathrm{d} \mathrm{a}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{g}\mathrm{e} \mathrm{c}\mathrm{o}\mathrm{l}\mathrm{o}\mathrm{r}=\sum {\mathrm{P}\mathrm{i}\mathrm{x}\mathrm{e}\mathrm{l}\mathrm{s}}_{\mathrm{c}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{g}\mathrm{o}\mathrm{r}\mathrm{y}}\times {\mathrm{C}\mathrm{o}\mathrm{l}\mathrm{o}\mathrm{r}}_{\mathrm{c}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{g}\mathrm{o}\mathrm{r}\mathrm{y}}$$

(1)

where Pixels_category is the fraction of pixels (0–1) in each category (dark green leaves, ribs, and inner leaves; light brown, medium brown, dark brown), and Color_category is the average color value (L*, a*, b*, or h°) of each category (dark green leaves, ribs, and inner leaves; light brown, medium brown, dark brown). The above formula allowed us to effectively obtain an average of the pixels of only the lettuce image after segmentation, avoiding any interference with the background. Finally, a browning index value was developed to represent the visual browning severity and combine all three browning categories from the segmentation (light, medium, dark) in terms of both the extent and the intensity of browning. This browning index defined a weighting for each category of brown pixels, with darker brown-colored pixels having higher importance:

$$\mathrm{B}\mathrm{I}={\mathrm{P}\mathrm{i}\mathrm{x}\mathrm{e}\mathrm{l}\mathrm{s}}_{\mathrm{\%} \mathrm{d}\mathrm{a}\mathrm{r}\mathrm{k} \mathrm{b}\mathrm{r}\mathrm{o}\mathrm{w}\mathrm{n}}\times 4+{\mathrm{P}\mathrm{i}\mathrm{x}\mathrm{e}\mathrm{l}\mathrm{s}}_{\mathrm{\%} \mathrm{m}\mathrm{e}\mathrm{d}\mathrm{i}\mathrm{u}\mathrm{m} \mathrm{b}\mathrm{r}\mathrm{o}\mathrm{w}\mathrm{n}}\times 2+{\mathrm{P}\mathrm{i}\mathrm{x}\mathrm{e}\mathrm{l}\mathrm{s}}_{\mathrm{\%} \mathrm{l}\mathrm{i}\mathrm{g}\mathrm{h}\mathrm{t} \mathrm{b}\mathrm{r}\mathrm{o}\mathrm{w}\mathrm{n}}\times 1$$

(2)

where Pixels_{% dark brown} is the percent of pixels in the dark brown category, Pixels_{% medium brown} is the percent of pixels in the medium brown category, and Pixels_{% light brown} is the percent of pixels in the light brown category [8].

2.4. Instrumental Quality Analyses

Headspace gas composition and electrolyte leakage were quantified for each lettuce sample. Headspace gas composition inside each package was measured 3, 7, and 14 days after processing. The oxygen and carbon dioxide concentrations were measured using a headspace gas analyzer (CheckMate II, PBI Dansensor, MOCON Inc., Minneapolis, MN, USA) by inserting a needle into the package through a gas-tight septum [14,15,17].

Electrolyte leakage was determined for the samples 7 and 14 days after processing using a method adapted from Luo et al. (2004) [14] in which lettuce samples (50 g) were combined with distilled water (500 mL) at 5 °C. After 30 min, the electrolyte content was quantified with a conductivity meter (Model 135A, Orion Research Inc., Beverly, MA, USA). The total electrolyte content was measured at 5 °C after the samples were exposed to three freeze–thaw cycles, where the samples were frozen to −20 °C for 24 h, followed by defrosting at room temperature (~22 °C). The electrolyte leakage after 30 min was calculated as a percentage of the total electrolytes after the three freeze–thaw cycles [14,15,17].

2.5. Statistical Analyses

Statistical analyses were conducted with SAS v9.4 (SAS Institute Inc., Cary, NC, USA) using an analysis of variance (ANOVA) with PROC GLM. Panelists were considered to be a random effect, and storage times and cultivars were analyzed as fixed effects for these data. Comparisons of treatment means were performed using Tukey–Kramer least significant difference tests, with significance defined as p < 0.05. Pearson correlation coefficients (PCCs) were calculated to determine the strength of the linear relationships between sensory and instrumental quality attributes; a PCC of >0.7 was interpreted as a strong correlation between variables, the 0.5–0.7 range indicated a moderate correlation, and <0.5 was considered a weak correlation. Principal component analysis was conducted to describe the relationship between attributes using latent factors that involved mean centering the data, disintegrating the data using eigenvalues, and nominating factors with the Scree test using PROC FACTOR in SAS v9.4 [29,49]. SigmaPlot v.13 (Systat Software, Inc., San Jose, CA, USA) was employed to graph the relationships between variables [40,50].

3. Results and Discussion

3.1. Sensory Evaluation

Sensory quality rating determined by panelists during storage time produced significant differences, particularly on days 7 and 14 after processing (Figure 4). Differences in late storage days of evaluations are common as imperfections become more visible once signs of deterioration due to higher respiration, cell lysis, and re-dox activity emerge [8,51]. The overall visual quality of all accessions was above 80 at day 3; however, acceptability decreased by day 7, and five of the accessions (Sun Valle, Green Forest, RH11-506, Tall Guzmaine, and Triple Threat) had an acceptability lower than 80. At day 14, all accessions had an acceptability of 60 to 80, with exception of PI 491224 and Triple Threat, which had lower acceptability scores (Figure 4A). Freshness scores showed a similar trend, with all accessions having an almost perfect score three days after processing, and a score of around 80 seven days after processing. The freshness scores for most accessions continued to decrease gradually to around 70, 14 days after processing; PI 491224 and Triple Threat had more rapid decreases at 14 days after processing (Figure 4B), with freshness scores of 60 and 20, respectively. The browning of tissue increased with the increasing number of days in storage. Three days after processing, all accessions had a browning score of less than 5. Seven days after processing, all accessions had scores of less than 16 for browning, except for Green Forest and RH11-1506, which had higher brown scores of 20 and 19, respectively. Fourteen days after processing, Sun Valley, Green Forest SM13-R2, Siskiyou, RH11-1506, Tall Guzmaine, Lobjoits, and King Henry had browning scores greater than 30, while the remaining accessions had browning scores of 10 to 30 (Figure 4C).

The above results suggest that panelists were not significantly influenced by the perception of browning tissues (at least below a certain undetermined threshold) when rating the overall visual quality of the samples. Freshness followed a similar trend as OVQ, indicating that these two traits are perceived similarly when assessing fresh-cut romaine lettuce. Another possible explanation is that the perception of freshness is a key determinant of a consumer’s evaluation of acceptability. Similar results have been observed when assessing diverse fresh produce [32,40], which have been linked more to the luminance distribution than to color [52,53].

The panelists’ scores for off odor at day 14 (Figure 5) showed that most of the accessions had an off-odor score below 5, except for King Henry (7), Tall Guzmaine (15), PI 491224 (21), and Triple Threat (74). These results matched the trends found in freshness and acceptability; PI 491224 and Triple Threat had the greatest off-odor scores, as well as the lowest freshness and OVQ scores.

The principal component analysis using the browning and OVQ sensory data (Figure 6) showed evidence of the opposite perception for these two factors, which was expected given that consumers typically have a negative association with browning in RTE lettuce. PI 491224 was located outside of the linear correlation curve form when pairing the two factors. Triple Threat fell even farther away from the trend line, implying that those two accessions differ significantly from all other tested accessions with respect to perception of OVQ and browning. It has been previously documented that both PI 491224 and Triple Threat are highly prone to a very rapid deterioration due to physiological decay that results in dark soft tissue [48], which often occurs before or at the same time as the common phenolic oxidation and browning of tissue. This common phenolic oxidation has been the main process observed for most other lettuce cultivars and breeding lines [8,29] and the focus in this study. The unusually high physiological decay, which also carries dark tissue, may have caused these two accessions to fall outside the linear trend. Additionally, sensory panelists were instructed to score browning of the lettuce samples based on the appearance of cut-edge browning only, excluding decay or decomposition. This explains why PI 491224 and Triple Threat showed lower browning scores while also a high level of decay.

3.2. Color and Browning Severity

When comparing the differences observed in incidence and severity of browning across accessions via CVS and image analysis (Figure 7), the accession with the least amount of browning was Clemente, with the lettuce pixels being identified as 69.4% regular pieces, 25.9% light brown, 3.3% medium brown, and 1.0% dark brown. On the opposite end of the spectrum, Triple Threat had the most browning, with the lettuce pixels having 53.6% regular pieces, 31.3% light brown, 7.0% medium brown, and 5.3% dark brown. The percent of light brown pixels was used to group the accessions into four phenotypic clusters: 1. Clemente (least browning group); 2. Parris Island Cos, Hearts Delight, Green Towers, Brave Heart, and PI 491224; 3. Sun Valley, Green Forest, RH11-1506, Lobjoits, SM13-R2, and Triple Threat; and 4. Siskiyou, Tall Guzmaine, and King Henry (most browning group). There was ≥3% difference between each group. When medium brown or dark brown pixel data were used for the analysis, only Triple Threat showed a uniquely high level of browning. Light browning is associated with physiological oxidation, and it is commonly observed in a dry state before decay begins [28]. On the other hand, dark browning is generally associated with the appearance of water-soaked tissue due to physiological deterioration and/or microbial decay [54,55], which was the process observed in Triple Threat in this and previous studies [8]. This observation indicates that the categorization of browning through image segmentation is an effective approach to quickly identify samples with a rapid rate of decay after processing.

When all the color categories were taken into consideration jointly through the empirically developed browning index, a clearer (statistical) differentiation was possible among some of the accessions. Tall Guzmaine, King Henry, and Triple Threat in general showed higher brown tissue than that in Clemente, Hearts Delight, and Parris Island Cos (Figure 8). Other differences within clusters were not observed due to data variability and subsequent standard deviations. However, the browning index was useful for providing a ranking of accessions that could be compared with the ranking produced by sensory panelists. With a few exceptions, a similar ranking of the accessions was produced using the browning index and visual evaluations by the panelists (Figure 4 and Figure 8).

3.3. Quality Parameters Determined by Instruments

No significant difference was observed among accessions in oxygen depletion, which was likely due to a significant effect of both MAP and refrigerated conditions in all samples. Likewise, the carbon dioxide content in the headspace of all 13 accessions was similar. The only exceptions were PI 491224 and Triple Threat, which had a significantly higher CO₂ concentration (Figure 9). The higher accumulation of CO₂ in samples of these two accessions could be associated with enzyme activity and phenolic composition [56], and/or microbial growth [57] commonly associated with soft decay and electrolyte leakage. Electrolyte leakage values showed similar results as the CO₂ analyses.

Electrolyte leakage measured 7 and 14 days after processing ranged from 0.7 to 1.5% for all accessions except Triple Threat, which had a significantly higher electrolyte leakage (12.4 to 14.1%) at both evaluation times, and PI 491224, which had higher electrolyte leakage (4.6%) at 14 days after processing (Figure 10). Both CO₂ concentration and electrolyte leakage values confirmed that these two cultivars are very highly prone to tissue deterioration due to cell lysis. The early stage of the deterioration process that is manifested as water-soaked, dark tissue areas was successfully captured by the image segmentation approach. The results of the CVS segmentation approach agree with those obtained using hyperspectral imaging and chlorophyll fluorescence imaging [58].

3.4. Relationships between Attributes

The main goal of the study was to determine whether a browning index derived from CVS could be used for a rapid assessment of the quality of fresh-cut lettuce. BI produced noteworthy correlations, with the key factors driving sensory and quality perception of panelists (and potentially also consumers). In particular, BI showed a strong correlation with freshness (r = −0.84), OVQ (r = −0.85) and off odor (r = 0.79) (Figure 11). Aside from our main objective of determining to what extent the empirical index could be of use, the variables measured with instrumentation (CO₂, O₂, electrolyte leakage) highly correlated with off odor, suggesting the importance of microbiological deterioration, as mentioned before.

There are many opportunities to apply machine learning and computer vision approaches to allow for non-destructive quality assessment of food products, including techniques such as support vector machine, K-nearest neighbor, decision trees, artificial neural networks, and convolutional neural networks [59]. The analysis of digital images as a quality evaluation tool can create new, more cost-effective alternatives to traditional sensory assessments of fresh produce in research and commercial settings, especially since techniques could be developed to work with images captured with cheaper point-and-shoot cameras or smartphones. While variables such as texture and aroma limit the use of this technology for some fruits, the power of images still appears to be significant to forecast in-situ preferences and eventually influence online consumers and e-grocery trends [60].

Our work adds to emerging reports that showed good performance of CVS-based models to predict quality in leafy vegetables without panelists’ in-situ intervention. For example, quality and deterioration of fresh-cut lettuce in clear plastic bags was well correlated using hyperspectral and chlorophyll fluorescence imaging approaches [58] and with a deep learning architecture based on a convolutional neural network [61]. Likewise, our study showed that browning quantified through image analysis was highly correlated to freshness, a subjective but critical factor in consumers’ decision-making at the point of purchase. The browning index developed in this study, or a modification of it, could be furthered in the future to predict overall product acceptability and freshness. This opens the possibility for fresh produce suppliers to virtually showcase a quality trait (freshness) associated with good postharvest handling techniques and timely coordination from field to retail and/or the local supply chain. Koyama et al. [42] also recently reported an effective model to predict consumers’ perception of the freshness of baby spinach. Their results assessing product freshness suggest that while using images for sensory analysis has limitations, it confirmed that machine learning and digital image analysis facilitate the assessment of a quality variable that is often influenced by light conditions in the room (i.e., lab or retail outlet). Novel apps that allow straightforward changes to a photograph background and improved technology to capture 360° and 3D images warrant further research to elucidate ways to maximize the use of electronic images in food-quality assessment. Such CVS-based approaches may be able to replace or limit the overreliance on panelists and instrumentation analysis to screen raw materials and track quality during postharvest handling processes.

4. Conclusions

This study demonstrated that the postharvest quality of romaine lettuce can be effectively evaluated using computer vision and image analysis techniques. We developed a CVS-based method that generated browning parameters that were highly correlated to key sensory attributes such as OVQ and freshness. The CVS results also agreed with results obtained with instrumentation (i.e., headspace gas composition and electrolyte leakage). Overall, the findings from this study illustrate a practical alternative to expedite screening for raw materials of RTE salads, which can be of potential application in the fresh-cut produce industry and breeding research programs.