Using Novel Optimized Deep Learning Techniques for Detecting Fungal Infections in Hazelnuts Kernels Based on Shell Color Changes

Abstract

Improper environmental or storage conditions can lead to fungal development in some hazelnuts. This type of fungus causes a slight discoloration on the hazelnut shell. To increase the marketability of the product, it is essential to separate these defective samples from sound ones. Due to the high color similarity between defective and sound samples, manual separation is prone to significant errors, necessitating an automated process. Furthermore, given the strong feature identification and extraction capabilities of convolutional neural networks (CNNs), this model was employed for the classification task. This study investigated the effect of some important factors such as input image size, flattening methods (global average pooling (GAP) and fully connected layer (FCL)), as well as the number of hidden layers and dropout on the model performance. In examining the effect of input image size on the models’ performance, the highest classification accuracy was obtained with a moderate image size of 128 × 128. Comparing the FCL and GAP methods indicated that the GAP method not only increased training speed but also minimized overfitting, resulting in overall better performance than FCL. The results of the models revealed that the proposed CNN model with four convolutional layers, employing the GAP method, a dropout rate of 0.5, and no hidden layers achieved the highest performance. The results demonstrate that the proposed CNN model effectively classified hazelnuts based on subtle color variations on their shells. It achieved a 0.8% improvement in detection accuracy compared to the manual classification method.