Instance segmentation method based on target contour

The purpose of instance segmentation is to use pixel-level instance mask to classify and locate a single object, which is a challenging task in computer vision. At present, there are some problems such as slow segmentation speed, low segmentation accuracy for small objects, and uneven segmentation edge. Aiming at the above problems, a instance segmentation method based on target contour is proposed. On the one hand, the feature extraction is carried out for the nodes on the object contour, which is not affected by the detection box and avoids processing the internal pixels of the object, so as to accelerate the segmentation speed. On the other hand, the gradual segmentation of the image is adopted to extract the features on the contour of the object multi-level, and the multi-scale fusion feature processing method is used to better extract the context semantic information and reduce the feature loss. The average segmentation accuracy of Cityscapes and KINS is 32.4% AP and 32.0% AP, respectively, which is better than the segmentation accuracy of many excellent works. The smoothness and the degree of fit of the object segmentation edge have better processing effect. The instance segmentation network based on target contour has better segmentation ability in instance segmentation task.