Research On Instance Segmentation Method Based On Anchor Free

Instance segmentation is one of the most basic and important research directions in computer vision.As a necessary step for image recognition from rough to fine,it can provide different instance tags for different individuals belonging to the same class.In recent years,the field of instance segmentation has made great progress,but there are still difficulties to be solved.The application of “anchor” will bring about many hyperparameters and complex post-processing.Starting from the perspective of “anchor-free”,this paper proposes an anchor-free instance segmentation method based on the residual aggregation network that can be trained end-to-end: ReAgMask.This paper also makes improvements and studies on the problems of low segmentation accuracy,imprecise segmentation of adjacent target edges,and slow model running speed.The research content of this paper is summarized as follows:(1)The anchor mechanism and existing instance segmentation algorithms are studied.This paper studies the representative instance segmentation algorithm and the anchor mechanism in them.After research and analysis,it is found that although the application of anchor points can improve the segmentation accuracy,it needs to set the hyperparameters according to the actual situation,and also requires complicated post-processing.(2)The residual aggregation network and instance segmentation algorithm based anchor-free(ReAgMask)are studied.This paper studies the advantages of Res Net and Dense Net and proposes residual aggregation networks based on task requirements.In order to solve the impact of anchor,this paper adds a Mask Head on the anchor-free detection algorithm,and constructs an instance segmentation framework based on anchor-free.Experimental results show that the segmentation accuracy,parameter amount and running speed of the algorithm proposed in this paper are better than the algorithms based on anchor,and the residual aggregation network performs better than Res Net in the algorithm.(3)The methods to improve the segmentation accuracy of ReAgMask are studied.Starting from the shortcomings of the algorithm,this paper introduces a new mask scoring mechanism to improve the quality of the generated mask.In order to finely segment adjacent target edges,this paper embeds the channel attention module in the backbone,and embeds the hybrid attention module in the Masd Head,so that the network can better focus on the target category and location information.In order to improve the segmentation accuracy of the network for objects with special shapes,this paper replaces the ordinary convolution in the residual aggregation network with deformable convolution to improve the feature extraction ability of the backbone network for objects with special shapes.The experimental results show that the improved model has a 1.7 point improvement in segmentation accuracy.(4)The methods to reduce the inference time of ReAgMask are studied.On the basis of the pre-work,this paper uses the mini version of the residual aggregation network after the channel is halved as the backbone,which greatly reduces the amount of network parameters without losing too much performance.In order to solve the problem that the bounding box quality estimation branch network is not easy to converge after the network is lightened,the classification quality joint loss function is used to monitor the quality of the generated bounding box.To reduce the time of NMS's inference time,this paper will use the Fast NMS.The experimental results show that the lightweight algorithm has less inference time and can achieve competitive segmentation accuracy.The inference time is reduced by 44.7ms,which is only two-fifths of the time before optimization.