Hybrid attention network for citrus disease identification

Accurate identification and timely prevention of citrus diseases will effectively protect the interests of the citrus industry. However, the citrus disease identification models currently used in the industry have unsatisfactory performance due to low robustness. In this study, we comprehensively study the problem of citrus disease identification from both data and algorithm perspectives. In order to address the root cause of the negative impact of the actual complex orchard environment on the identification model in practical applications, an orchard context-based citrus disease dataset including Citrus yellow vein clearing virus (CYVCV), Canker, Brown spot, Melanose, Sooty mold, and healthy control is created. Data are collected from citrus leaf parts to maintain data uniformity. The model trained in this dataset can better adapt to the complex environment of the orchard, and thus more effectively serve the application in the production area. The key information for citrus disease identification is spot characteristic information, but due to the small size of the spots, it is difficult to focus and extract the characteristic information. In order to solve this problem, we studied that the representation of the features in the frequency domain dimension after the wavelet transform process is sparse, which is beneficial to improving the performance of the attention module, and proposed the frequency-domain attention network (FdaNet) to adaptively learn through the importance of feature information between different frequency domains changes the weight of each frequency domain during network inference. The effectiveness of FdaNet was demonstrated in experiments on citrus disease identification embedded in a ResNet backbone network. Next, according to the complex and diverse background of citrus disease data, a hybrid attention network (HaNet) is proposed to focus on multi-dimensional feature information. In HaNet, the frequency domain attention module is embedded into the channel attention network to enhance the channel scalar computing capability. In addition, in order to maximize the feature range extracted by the attention module in two dimensions, large convolution kernels are introduced in the backbone network to improve the effective perceptual field of the network. Moreover, we conducted research experiments using large convolution kernels of different sizes to further select sensory fields suitable for citrus disease feature extraction. Experimental results on the citrus disease dataset show that our proposed model achieves recognition accuracy of 98.83 % and 98.77 % on 50-layer and 101-layer networks respectively, both of which are better than other state-of-the-art models.