A pooling module with multidirectional and multi-scale spatial information and its application on semantic segmentation of leaf lesions

Timely and accurate identification of apple leaf diseases provides an important basis for the early warning and precise control of apple leaf diseases. It was of great significance to reduce the economic losses caused by diseases. In order to improve the accuracy of apple leaf lesion segmentation by using the deep neural networks, this paper proposed a twill pooling method and combined it with strip pooling to propose the double cross pooling method. The pooling method could contract multidirectional and multi-scale spatial context information. Then, on the constructed apple leaf disease dataset, these modules were combined with existing deep semantic segmentation networks to perform disease lesion semantic segmentation. These three constructed modules were added to the fully convolutional network respectively, and the universality of the constructed modules was further verified on other semantic segmentation networks. Experimental results showed that the proposed modules could improve the mean intersection over union of fully convolutional network by 7.69%, up to 80.44% on the collected apple leaf disease dataset. Furthermore, when adding the proposed pooling modules to DeepLabV3 + , PSPNet and U-Net, the mean intersection over union improved by varying degrees when adding any of the proposed modules alone. The performance of the improved fully convolutional network, DeepLabV3 + , PSPNet and U-Net were compared. The DeepLabV3 + with the double cross pooling module had the best segmentation performance whose mean pixel accuracy, mean intersection over union, leaf disease classification accuracy and leaf disease degree diagnosis accuracy were 99.07%, 82.1%, 99.52% and 77.65% respectively.