Semantic Segmentation of Weeds Based on Multiscale Information Fusion
Effective segmentation of weeds in maize fields is a prerequisite for accurate variable herbicide application by UAVs.Aiming at the problems of missing detection of the traditional CNN semantic segmentation model for the mutual occlusion of maize and weeds,and the small target weeds,etc.,taking UAV ortho-digital image of the field weed at three-five-leaf-age of maize as the object,we propose the weed semantic segmentation model TFPSP-CA based on multi-scale information fusion of Transformer and CNN.Firstly,the base model PSPNet feature pyramid is replaced with BiFPN to strengthen feature fusion and enhance the model's ability to learn image details and acquire contextual information;the original ResNet network is replaced with the MobileNet series network to speed up the model prediction speed and reduce the model size,and the improved weed segmentation model FPSPNet is obtained.The results show that the mloU and PA of the improved segmentation model FPSPNet are 84.48%,89.36%,82.05%,89.34%,and 83.32%,89.26%for small,medium,and large targets,respectively.The accuracy of weed segmentation is especially improved significantly for small targets,and its mIoU and PA are increased by 4.60%and 2.42%respectively compared to the base segmentation model.Secondly,the Transformer feature output module is further introduced at the output end of the FPSPNet pyramid module to connect in parallel to obtain more scale information.The CA attention mechanism combining coordinate and channel information is added to obtain the TFPSP-CA weed segmentation model.The model can better capture the global features and long-distance dependence,thus improving the processing ability of the weed segmentation problem in complex environments.The results show that the improved TFPSP-CA model has 90.21%,91.98%,89.44%,89.11%,and 87.59%,87.53%of weed segmentation mIoU and PA in the case of no occlusion,mild occlusion,and severe occlusion,respectively,and the accuracy of the improved model in the case of severe occlusion improves significantly,compared with the original models PSPNet and FPSPNet,the mIoU and PA are improved by 6.34%,2.27%and 10.96%and 5.22%,respectively.
国家科技期刊平台
NSTL
万方数据
