基于选择性注意力神经网络的木薯叶病害检测算法

为了实现在复杂非结构环境下对木薯叶4种主要病害的高精度检测，提出一种基于选择性注意力机制的木薯叶病害神经网络检测改进算法MAISNet (Multiattention IBN Squareplus neural network)。以V2-ResNet-101为基础网络，先使用多重注意力算法优化加权系数，调整特征通道的语义表达，在特征图中初步构建显著性特征；然后在残差单元之后采用实例批归一化方法来抑制特征表达中的协变量偏移，在特征图中构建出显著性语义特征，实现高质量语义特征表达；最后在残差分支中采用Squareplus激活函数替代ReLU激活函数，保持语义特征在负数域的数值分布，减少特征拟合过程中的截断误差。对比试验结果显示，经过上述改进后构建出的MAISNet-101神经网络，对4种常见木薯叶病害检测的平均准确率达到95.39%，明显优于目前主流算法EfficientNet-B5和RepVGG-B3g4等。网络提取特征的可视化分析结果表明，高质量木薯叶病害显著性语义特征，是提高木薯叶病害检测准确率的关键。所提出的MAISNet神经网络模型可以完成实际场景下木薯叶病害高精度检测。

Cassava Leaf Disease Detection Algorithm Based on Selective Attention Neural Network

To achieve high-precision detection of four major cassava leaf diseases in complex unstructured environments, an improved algorithm for cassava leaf disease neural network detection based on the selective attention mechanism, MAISNet, was proposed. Using V2-ResNet-101 as the base network, the multiattention algorithm was firstly used to optimize the weighting coefficients, adjust the semantic expression of the feature channels, and the semantic feature saliency expression of cassava leaf disease in the feature map was preliminary constructed; then the instance batch normalization method was used after the residual unit to suppress the covariate offset in the feature expression, highlight the target semantic feature expression in the feature map, and realize the high-quality semantic feature expression. Finally, the Squareplus activation function was used to replace the ReLU activation function in the residual branch to maintain the numerical distribution of semantic features in the negative domain, and reduce the truncation errors in the feature fitting process. The results of the comparison test showed that the MAISNet-101 neural network constructed after the above improvement achieved an average accuracy of 95.39% for the detection of four common cassava leaf diseases, which was significantly better than the performance of the mainstream algorithms such as EfficientNet-B5 and RepVGG-B3g4. The results of the visualization and analysis of the extracted features of the network showed that high-quality semantic feature saliency representation of cassava leaf diseases was the key to improve the accuracy of cassava leaf disease detection. The proposed MAISNet neural network model can accomplish high-precision detection of cassava leaf diseases in real scenarios, which can provide technical support for precise drug application.