Hierarchical multi-scale feature fusion network based for apple leaf disease recognition

Due to the small inter class differences and large intra class differences in apple leaf disease images, effectively integrating global and local multi-scale features is the key to apple leaf disease recognition. This article proposes a hybrid CNN Transformer architecture called HC-Former for apple leaf disease recognition. HC-Former consists of three main modules: CNN encoder, Transformer encoder based on state space model (VSSM), and cross attention fusion module. The CNN encoder effectively restores local spatial information and improves fine scale details. Meanwhile, the VSSM encoder captures remote dependencies between sequences of arbitrary length, allowing them to obtain global features and contextual information. The cross attention fusion module integrates local and global features, and the advantage of this network structure is that the generated representation semantics are richer, which can effectively extract local features and global representations at different semantic scales. The comparative experiment results based on standard image set and real image set show that after layer by layer feature fusion of CNN encoder and Transformer encoder based on state space model, the recognition accuracy of HC-Former model reaches 97.65% and 98.16% respectively, which is 4.19% and 3.92% higher than the baseline model, and can meet the application requirements of ALD recognition in smart agriculture.