Research On Image Instance Segmentation Based On Multi-level Feature Fusion

In recent years,with the development of deep learning theory and technology,related research in the field of computer vision has become more and more intensive.In the field of image segmentation,two branches have been refined: semantic segmentation and instance segmentation.The task of instance segmentation is to use a computer to extract objects of human interest and shield out noise information such as background.This paper summarizes the current situation at home and abroad for the task of case segmentation,and studies the key issues in case segmentation.The main tasks include:(1)Image instance segmentation network based on multi-level feature fusionThe example segmentation algorithm Mask-RCNN's semantic segmentation branch network uses a simple full convolutional network with limited performance.This paper improves the branch network,and proposes a feature fusion module and a pooling chain module.Through the processing of these two modules,not only the performance of network extraction features can be enhanced,image context information can be obtained in a large background,but also the loss of spatial information caused by convolution and pooling processing can be avoided.Experiments on chromosome datasets and plant datasets show that the improved algorithm in this paper has a significant improvement in segmentation performance.(2)Image instance segmentation network combining self-attention mechanism and path enhancementThe traditional feature pyramid fusion network uses simple feature addition.For the highest-level feature maps,there is no deep mining of effective information.The fusion method is only performed in two-dimensional space.In this regard,this paper proposes a self-attention module to mine the highest-level feature maps in depth,extract multi-scale information from them and fuse them.The path enhancement structure is further designed,and the spatial dimension attention is introduced on the premise of the original feature pyramid structure.The low-level feature map is weighted to guide the direction of feature fusion.Experimental results verify the effectiveness of the method.