Apple Leaf Disease Identification Through Region-of-Interest-Aware Deep Convolutional Neural Network

Traditional approaches for the identification of leaf diseases involve the use of handcrafted features such as colors and textures for feature extraction. Therefore, these approaches may have limitations in extracting abundant and discriminative features. Although deep learning approaches have been recently introduced to overcome the shortcomings of traditional approaches, existing deep learning models such as VGG and ResNet have been used in these approaches. This indicates that the approach can be further improved to increase the discriminative power because the spatial attention mechanism to predict the background and spot areas (i.e., local areas with leaf diseases) has not been considered. Therefore, a new deep learning architecture, which is hereafter referred to as region-of-interest-aware deep convolutional neural network (ROI-aware DCNN), is proposed to make deep features more discriminative and increase classification performance. The primary idea is that leaf disease symptoms appear in leaf area, whereas the background region does not contain useful information regarding leaf diseases. To realize this, two subnetworks are designed. One subnetwork is the ROI subnetwork to provide more discriminative features from the background, leaf areas, and spot areas in the feature map. The other subnetwork is the classification subnetwork to increase the classification accuracy. To train the ROI-aware DCNN, the ROI subnetwork is first learned with a new image set containing the ground truth images where the background, leaf area, and spot area are divided. Subsequently, the entire network is trained in an end-to-end manner to connect the ROI subnetwork with the classification subnetwork through a concatenation layer. The experimental results confirm that the proposed ROI-aware DCNN can increase the discriminative power by predicting the areas in the feature map that are more important for leaf diseases identification. The results prove that the proposed method surpasses conventional state-of-the-art methods such as VGG, ResNet, SqueezeNet, bilinear model, and multiscale-based deep feature extraction and pooling.