Novel models to predict stored melon fruit marketability using convolutional neural networks.

This research aimed to classify melon fruit of the cultivar TM-A94 (Cucumis melo var cantalupensis) into marketable, suitable to produce fresh-cut fruit, or storage loss by employing convolutional neural network models. For this purpose, melon fruit were harvested and stored at three different temperatures (1, 4, and 12 °C). Fruit were evaluated for marketability, inner qualities, and taste at harvest and throughout cold storage. Images of all evaluated fruit were acquired, labeled according to the proposed classification, and preprocessed. The proposed models consisted of consecutive basic convolutional modules and different input sizes (224 × 224 × 3, 112 × 112 × 3, and 56 × 56 × 3). Furthermore, SqueezeNet classifier was evaluated using the transfer learning approach. Different learning rates, optimizers, and weight decay levels were tested with the studied models to choose the optimum training settings. Augmentation techniques were used to expand the acquired dataset. fivefold cross-validation was used to validate the models through optimization and training. The final validation accuracies were 90, 90, 86, and 86.33% for the trained SqueezeNet, 224p, 112p, and 56p models, respectively. Confusion matrices showed that SqueezeNet and 224p could detect the marketable class sufficiently, with an F1-score reaching 94.67 and 92.98%, respectively. However, the highest mispredictions among all models were observed between the marketable class and the class suitable to produce fresh-cut. Additionally, the results suggest using high-resolution images (224 × 224 or more) for similar classification tasks in melon fruit. [ABSTRACT FROM AUTHOR]