Improving the accuracy of real field pomegranate fruit diseases detection and visualisation using convolution neural networks and grad-CAM

Pomegranate (Punica granatum L) is one of the vital cash fruit crops of arid and semiarid regions in India. The occurrence of pests and diseases affects the development and quality of fruits. Our objective is to develop an automated pomegranate disease detection system on an actual field image dataset using convolution neural networks. The collected images are classified into six categories namely healthy, bacterial blight, anthracnose, fruit spot, fusarium wilt, and fruit borer. In this paper, we have measured the performance of CNN-based architectures VGG16, VGG19, InceptionV3, Resnet50, and Xception with hyperparameter tuning. The experimental results show that Resnet50 is a suitable model for our dataset with a disease detection accuracy of 98.55%. To deal with DL 'black box' problem, the gradient-weighted class activation mapping (Grad-CAM) model is integrated with ResNet50 to highlight the important regions on the fruits to locate accurate diseases and recommend appropriate disease treatment to farmers.