Crop image segmentation method based on improved Mask RCNN

In the image of agricultural products, due to the wide variety and the surrounding environment information is easy to be confused and segmented, it is difficult for traditional methods to achieve the problem of accurate and efficient extraction of crops. This paper takes Mask RCNN as the research object and uses the PyTorch deep learning framework to focus on the research and construction of a network structure based on the improved crop detection segmentation of Mask RCNN. ResNet50 and FPN are combined as the backbone network for feature extraction and target candidate regions are generated. The Softer-NMS algorithm is added to the extraction network (RPN) to remove redundant anchor frames, improve the accuracy of crop detection and positioning, and optimize the feature pyramid network (FPN) structure and integrate multi-scale feature maps to improve the accuracy of instance segmentation. The spatial information of the feature map is saved by the bilinear interpolation method in ROIAlign. Compared with the image extraction algorithms such as Faster RCNN, Faster RCNN with FPN and Mask RCNN, the experimental results show that the improved Mask RCNN algorithm proposed in this paper has good average accuracy, detection and masking in crop image extraction. The precision of the extraction results, Recall, Average precision and Mean Average Precision (mAP) score indicators are better than the comparison algorithm.