基于非对称混洗卷积神经网络的苹果叶部病害分割

针对苹果叶部病害由于数据集类间样本不均衡和拍摄角度、光照变化等实际成像与环境因素造成的精度低和泛化能力差的问题，本文提出了一种新型的非对称混洗卷积神经网络ASNet。首先，通过在ResNeXt骨干网络中添加改进的scSE注意力机制模块增强网络提取的特征;其次,针对多数叶片病害特征分布相对分散的问题，使用非对称混洗卷积模块代替原始的残差模块来扩大卷积核的感受野和增强特征提取能力，从而提升模型的分割精度和泛化能力;最后,在非对称混洗卷积模块中使用通道压缩和通道混洗的方式弥补了分组卷积造成的通道间关联性不足的缺陷，降低了由于叶部病害类间不均衡导致的传统网络模型精度偏低的问题。在COCO数据集评价指标下，实验结果表明，相比于骨干网络为ResNeXt-50的原始Mask R-CNN模型，本文模型的平均分割精度达到96.8%，提升了5.2个百分点，模型权重文件减小为321MB，减小了170MB。对实地采集和AI Challanger农作物病害分割挑战赛的240幅苹果叶片图像进行测试，结果表明，本文模型ASNet对苹果黑腐病、锈病与黑星病3种病害和健康叶片的平均分割精度达到94.7%。

High Precision Identification of Apple Leaf Diseases Based on Asymmetric Shuffle Convolution

Aiming at the problems of low accuracy and poor generalization ability caused by the imbalance of samples between data sets, shooting angles, light changes and other actual imaging and environmental factors caused by apple leaf diseases, a type of asymmetric shuffle convolution neural network ASNet was proposed. Firstly, by adding an improved scSE attention mechanism module to the ResNeXt backbone network to enhance the network feature extraction; secondly, for the relatively scattered feature distribution of most leaf diseases, the asymmetric shuffle convolution module was used to replace the original residual module to expand the receptive field of the convolution kernel and the enhanced feature extraction ability, thereby improving the recognition accuracy and generalization ability of the model; finally, the use of channel squeeze and channel shuffling in the asymmetric shuffle convolution module made up for the grouping convolution. The defect of insufficient correlation between channels reduced the problem of low recognition accuracy of traditional network models caused by the imbalance between leaf diseases. Under the COCO data set evaluation index, the experimental results showed that compared with the Mask R-CNN whose backbone network was ResNeXt-50, the average test accuracy of this model reached 96.8%, which was increased by 5.2 percentage points, and the model size was reduced to 321 MB, a decrease of 170 MB. Tested by 240 field-collected and AI Challanger crop disease identification challenge apple leaf images, the test results showed that the average segmentation accuracy of the proposed model ASNet for apple black rot, rust, scab and healthy leaves reached 94.7%.