Research And Improvement Of Instance Segmentation Algorithm Based On Deep Learning

In recent years,the development of deep convolutional neural network is very rapid.The prediction ability of neural network is gradually developing from image level prediction to region prediction,pixel prediction and instance prediction.Instance segmentation is both practical and challenging,and has become a research hotspot in computer vision.Mask RCNN is a simple and effective framework for instance segmentation,which plays an important role in the field of instance segmentation due to its good effect and easy expansion.However,it still has some problems such as false detection,missed detection,and low instance segmentation accuracy.To solve these problems,this paper improved the Mask RCNN algorithm.In addition,in order to solve the problem of instance segmentation under specific scenarios,the improved instance segmentation algorithm is also discussed on the selfmade dataset.First of all,in order to improve the accuracy of Mask R-CNN algorithm and improve the missed and false detection,this paper makes three improvements on feature extraction and fusion.First,a deformable convolution layer is used in the feature extraction layer to expand the perceptive field,make the perceptive field more suitable to the shape of the target,and improve the precision of instance segmentation.Secondly,a spatial and channel excitation module is proposed to improve the missing and error detection in the instance segmentation network.Thirdly,by using an inverted FPN network,the details of the bottom layer are quickly introduced into the top layer by adding very few convolution layers,so that the features of the bottom layer can be easily propagated.The experimental results show that both detection and instance segmentation have been improved significantly,and both of them have been improved on the public MS COCO dataset.In addition,this paper makes a dataset of indoor human body for instance segmentation.There are serious occlusion and multi-scale problems among the humanoid objects in the dataset,which leads to the poor segmentation effect of the instance.Therefore,this paper makes some improvements on the characteristics of the self-made dataset.First,Soft-NMS is used to improve the condition of missed detection caused by the setting of NMS threshold.Secondly,in view of some characteristics of self-made datasets,two kinds of public datasets are used in this paper to conduct directional expansion to enhance the robustness of the network,and good results are obtained on the self-made dataset.