Automated Tomato Defect Detection Using CNN Feature Fusion for Enhanced Classification

Tomatoes are among the most widely cultivated and consumed vegetable crops worldwide. They are usually harvested in large quantities that need to be promptly and accurately classified into healthy and defective categories. Traditional methods for tomato classification are labor-intensive and prone to human error. Therefore, this study proposes an approach that leverages feature fusion from two pre-trained convolutional neural networks (CNNs), VGG16 and ResNet-50, to enhance classification performance. A comprehensive evaluation of multiple individual and hybrid classifiers was conducted on a dataset of 43,843 tomato images, which is heavily imbalanced toward the healthy class. The results showed that the best-performing classifier on fused features achieved an average precision (AP) and accuracy of 0.92 and 0.97, respectively, on the test set. In addition, the experimental evaluation revealed that fused features improved classification performance across multiple metrics, including accuracy, AP, recall, and F1-score, compared to individual features of VGG16 and ResNet-50. Furthermore, the proposed approach was benchmarked against three standalone CNN models, namely MobileNetV2, EfficientNetB0, and DenseNet121, and demonstrated superior performance in all evaluated metrics. These findings highlight the efficacy of deep feature fusion in addressing class imbalance and improving automated tomato defect detection.