Leaf Spot Attention Network for Apple Leaf Disease Identification

Abstract

Although new deep learning approaches have recently been introduced for leaf disease identification, existing deep learning models such as VGG and ResNet have been used previously. Therefore, a new deep learning architecture is proposed to consider the leaf spot attention mechanism. The primary idea is that leaf disease symptoms appear in the leaf area, whereas the background region does not contain any useful information regarding leaf diseases. To realize this, two subnetworks are designed. The first is a feature segmentation subnetwork to provide more discriminative features for the separated background, leaf areas, and spot areas in the feature map. The other is a spot-aware classification subnetwork to increase the classification accuracy. To train the proposed leaf spot attention network, the feature segmentation subnetwork is first learned with a new image set, where the background, leaf area, and spot area are annotated. Subsequently, the spot-aware classification subnetwork is connected to the feature segmentation subnetwork and then trained through early and later fusions to produce the semantic-level spot feature information. The experimental results confirm that the proposed network can increase the discriminative power by modeling the leaf spot attention mechanism. The results prove that the proposed method outperforms conventional state-of-the-art deep learning models.