Segmentation of Marine Targets Instance Based on Improved Mask R-CNN

Instance segmentation is the classification and segmentation of targets in images by pixel-level masks. The method for instance segmentation, Mask R-CNN, performs well in the task of segmenting targets such as pedestrians and vehicles. However, it still has the problem of low accuracy for small targets and marine targets segmentation with complex contours, and information loss occurs in the process of feature extraction and fusion of images. In this paper, an improved Mask R-CNN algorithm is proposed. Feature pyramid network (FPN) is enhanced, and the attention mechanism is added between the backbone and the feature pyramid network to solve the problem of the loss of low-level semantic information in the process of feature fusion. The research involves constructing a marine targets instance segmentation dataset, learning on the dataset, and comparing the improved algorithm with the original algorithm by coco index. The results show that the improved algorithm has obvious improvement in accuracy, has strong robustness.