An Improved MobileNet Model Integrated Spatial and Channel Attention Mechanisms for Tea Disease

Abstract

Aiming at addressing the challenges of large model parameters, high computational cost, and low accuracy of the traditional tea disease identification model, an improved MobileNet model integrated spatial and channel attention mechanisms (MobileNet-SCA) was proposed for tea disease identification. Firstly, the tea disease identification dataset was augmented through random clipping, rotation transformation, and perspective transformation to simulate diverse image acquisition perspectives and mitigate overfitting effects. Secondly, based on the convolutional neural network (CNN) framework, the Channel Attention (CA) mechanism and Spatial Attention (SA) mechanism were introduced to carry out global average pooling and group normalization operations on input feature maps respectively, and adjust the channel weights using the learned parameters. Then the h-swish activation function was utilized to scale, and the two kinds of attention mechanisms were spliced and mixed to improve the channel and spatial information. In addition, the MobileNetV3 network's structure underwent optimization by adjusting the number of input channels, the size of the convolution kernel, and the number of channels in the residual block. The experimental results showed that the identification accuracy of MobileNet-SCA for tea diseases was 5.39% higher than the original model. This method can balance the identification accuracy and identification time well, and it meets the requirements for accurate and rapid identification of tea diseases.