Vision-based strawberry classification using generalized and robust deep networks

Grading of agricultural products such as fruits and vegetables based on ripeness level and visual defects for the purpose of export, storage and waste control is a process of special importance. Various methods have been used to detect levels of ripeness and the quality of agricultural products, some of which are destructive and some non-destructive. The machine vision system is one of the non-destructive and accurate systems in the field of detecting the quality of agricultural products. In this study, we propose a robust and generalized model via fine-tuning the pre-trained networks for the classification of strawberry fruit. A dataset containing 800 confirmed strawberry images in four classes (unripe, half-ripe, ripe, and damaged) was used. Instead of using fundamental data augmentation (FDA) techniques to prevent overfitting problem and increase the robustness of the model, we employed a novel learning-to-augment strategy (LAS) using noisy images that creates new noisy variant of data via original images. By using the Bayesian optimization algorithm, controllers were used to select the optimal noise parameters of Gaussian and speckle noise to generate new noise images. The best policies of data augmentation based on LAS was used to fine-tune pre-trained cutting-edge models (GoogleNet, ResNet18, and ShuffleNet). The results show that in all the proposed scenarios (i.e. using original data without data augmentation, employing FDA, and applying LAS) the GoogleNet model was able to achieve 96.88 %, 97.50 %, and 98.85 % accuracy, respectively.