Research On Image Instance Segmentation Algorithm Based On Multi-scale Feature Fusion

The instance segmentation task is one of the key research tasks in the field of computer vision.The main goal is to distinguish and recognize images from the pixel level,specify category labels,and distinguish different individuals in the scene.According to different input data,instance segmentation techniques can be divided into 2D image instance segmentation and 3D point cloud instance segmentation.With the continuous advancement of deep learning technology and convolutional neural networks,the task of 2D image instance segmentation has achieved good development,and attempts have been made to apply it to various tasks in the real world.However,in practical applications,the 2D image instance segmentation method needs to have a higher processing speed,and the processing speed of the current mainstream algorithms cannot meet real-time requirements.How to improve the processing speed of the 2D image instance segmentation algorithm is one of the research contents of this paper.The development of deep learning has also promoted the development of 3D scene understanding tasks.As one of the representation formats of 3D scene information,3D point cloud data has gradually become the main data representation of 3D scene understanding tasks due to its low acquisition cost and simple representation methods.method.At the same time,some researchers began to think about how to learn from the successful experience in the field of 2D image instance segmentation to perform instance segmentation on 3D point clouds.However,point cloud data is sparse,and direct use of dense convolution operations in 2D image processing will inevitably result in a serious waste of resources.How to design a good feature extraction scheme,extract features from 3D point cloud scenes,and complete instance segmentation is another research content of this paper.This paper mainly studies the instance segmentation method of 2D image and 3D point cloud and designs a more excellent instance segmentation task algorithm for the above problems.The research content of this paper is as follows:(1)In the 2D image instance segmentation task,because of the slow processing speed of previous algorithms and the difficulty in real-time processing of scene information,this paper proposes a single-stage real-time instance segmentation algorithm for 2D images based on anchor points,which extracts The feature of the 2D input image is obtained by using the feature pyramid algorithm to obtain the multi-scale feature map.Finally,the feature is directly processed by the simple mask prediction branch,and the final instance segmentation result is obtained,which improves the processing speed of the instance segmentation algorithm.(2)In the 3D point cloud instance segmentation task,the point cloud is usually disordered and sparse,and it is difficult to directly use the relevant processing ideas of the 2D image convolution scheme.Aiming at the characteristics of point cloud disorder,this paper proposes a 3D point cloud semantic instance segmentation algorithm based on a multi-scale feature extraction network,and constructs a feature extraction network in the form of an "encoder-decoder",so that it can Extract and encode local feature information,and improve the representation ability of the point cloud feature extraction network through multi-scale fusion operations;for segmentation tasks,the algorithm divides feature processing into two parallel subtasks for generating semantic features and instance embeddings respectively,and integrate the information of semantic features and instance embedding to improve the accuracy of instance segmentation.(3)In the 3D point cloud instance segmentation task,the point cloud scene to be processed is usually relatively large,and the point-based method uses the distance between point pairs as the measurement criterion.When the number of scene points increases,such methods will significantly increase the computational overhead.;Due to the sparsity of the point cloud,the traditional dense convolution method will gradually lose the sparse geometric features of the object,so it is necessary to design a better feature extraction scheme.Because of the above problems,this paper proposes a 3D point cloud instance segmentation algorithm based on sparse convolution and proposal generation.With the help of the sparse convolution method,a multi-scale feature extraction network is constructed to obtain input semantic prediction scores and offset vectors;to improve instance segmentation The performance of the algorithm can reduce misjudgment.The algorithm uses the semantic prediction score and offset vector of the previous step to generate instance proposals,and then inputs the instance proposals into another small feature processing network,and finally outputs accurate instance labels.