Tea disease detection method using fusion of channel and spatial attentions

Accurate and rapid detection of tea disease is of great economic significance to the tea industry. However, the problems such as scale variation, mutual occlusion and complex background of disease-affected leaves greatly reduce the prediction accuracy of the detector. A tea disease target detection method, integrating channel-spatial attention, was proposed to achieve accurate and efficient disease detection. Firstly, the multi-layer self-attention mechanism was used to extract multi-scale features of tea diseases to obtain local detail features of tea images at different scales. Secondly, the new channel attention module was introduced to extract richer channel information, suppress complex background noise and enhance the ability of model feature representation. In addition, a new spatial attention module was also proposed to further extract the spatial relationship of features, reduce redundant information and optimize computational overhead. The experimental results showed that the tea pest detection method with channel-spatial attention could cope with the challenges of leaf scale changes, mutual occlusion and complex background.