Abstract:  In order to solve the problem that traditional neural networks cannot meet the recognition needs of mobile devices with limited resources for apple leaf diseases due to the large number of parameters, a multi‑scale lightweight network model CA—MobileNetV2 based on coordinate attention was proposed. Firstly, the 3×3 convolution in the MobileNetV2 reciprocal residual is replaced with a multi‑scale feature fusion module (MMF—module). Without increasing the number of parameters, empty convolution is introduced to increase the sensitivity field, so as to capture rich multi‑scale details and enhance the ability of the network to extract details and semantic information. Secondly, coordinate attention mechanism is introduced to learn the feature weights of different positions adaptively to enhance the perception ability of apple leaf disease region. Finally, to solve the problem of gradient disappearance in model training, the MobileNetV2 classifier is improved and the Leaky ReLU activation function is introduced. The experiment result shows that the recognition accuracy, parameters, and FLOPs of the lightweight model proposed in this article on the apple leaf disease dataset are 98.36%, 2.35 MB, and 298.70 M respectively. Compared with ShuffleNetV2, EfficientNet—B2, MobileNetV2, MobileNetV3, and GhostNet, the parameter count has been compressed by 0.69 MB, 6.41 MB, 0.28 MB, 4.32 MB, and 1.46 MB, and the accuracy has been improved by 8.6%, 6.47%, 5.07%, 4.28%, and 3.85%, while the inference time has been reduced by 8.7 ms, 21.1 ms, 13 ms, 6.9 ms, and 17.6 ms.

Key words:  , apple leaves, disease identification, coordinate attention, lightweight model, muti?scale fusion features

摘要： 为解决传统神经网络参数量大、无法满足资源有限的移动设备对苹果叶片病害的识别需求，提出一种基于坐标注意力的多尺度轻量级模型CA—MobileNetV2。首先，将MobileNetV2倒残差中3×3的卷积替换成多尺度特征融合模块(MMF—module)，在不增加参数量的前提下，引入空洞卷积增大感受野，以捕捉丰富的多尺度细节信息，增强网络对细节信息和语义信息提取能力；其次，引入坐标注意力机制自适应地学习不同位置的特征权重，增强对苹果叶片病害区域的感知能力；最后，针对模型训练中的梯度消失问题，改进MobileNetV2的分类器，并引入Leaky ReLU激活函数。结果表明，所提轻量级模型在验证集上的识别准确率、参数量、浮点运算量分别为98.36%，2.35 MB和298.70 M，与ShuffleNetV2、EfficientNet—B2、MobileNetV2、MobileNetV3和GhostNet相比，参数量压缩0.69 MB、6.41 MB、0.28 MB、4.32 MB、1.46 MB，准确率提升8.6%，6.47%，5.07%，4.28%和3.85%，推理时间减少8.7 ms、21.1 ms、13 ms、6.9 ms、17.6 ms。

关键词: 苹果叶片, 病害识别, 坐标注意力, 轻量级模型, 多尺度特征融合