Improved Image Instance Segmentation Algorithm Based On Mask R-CNN

Image instance segmentation is a pixel level instance target segmentation task,which is an important development direction in the field of artificial intelligence and computer vision.It is widely used in various fields,such as industrial production,monitoring security and medical and health.There are two problems in the current image instance segmentation model.Firstly,the traditional image instance segmentation model has many problems such as false detection and missed detection in the detection stage because of the large difference in the size of the object in the image,which makes the mask precision lower.Secondly,the traditional instance segmentation model is used to classify and determine the pixel points by gathering the feature information of the whole connection layer in mask generation stage.Pool operation will lead to the decrease of feature map size.In the semantic segmentation stage,the feature information passed to the whole connection layer is lost,which leads to the deviation of the category belonging of pixel points,low mask quality and the detail information of the edge of the instance is not refined.In this paper,based on the existing instance segmentation algorithm framework mask r-cnn,we make in-depth research and improvement to solve the problems of large size span of object instance,low quality and incomplete mask generated in image instance segmentation task.The main contents of this paper are as follows:(1)In order to eliminate the influence of large size difference between instance objects,this paper designs a feature extraction network based on hierarchical multi-scale attention(HMSA),which is used to fuse the feature information of instance objects of different scales in the feature extraction process to improve the accuracy of instance segmentation.The network can learn which scale is preferred in a specific situation,and learn all the relative attention feature information between adjacent scale sets.The feature matrix obtained by the feature extraction network not only integrates multi-scale features,but also has preference characteristics.(2)In order to break the limitation between convolutional neural network and full connection layer and solve the problem that the ratio of width to height of validation data and training data is different.Aiming at the inaccurate pooling result of original spatial pyramid,an optimized spatial pyramid pooling method is proposed in this paper,which makes convolution neural network adapt to multi-scale and different proportion input.In addition,deconvolution operation is used for up sampling in the top-down path of feature pyramid to connect local features with different scale space under the action of transverse connection structure Feature stitching.(3)In the mask generation stage,the instance features are combined with the corresponding prediction mask to trace the mask IOU and correct the deviation between mask score and mask quality.Specifically,the semantic features extracted from the mask generation network and the instance class features in the feature extraction network are multi-level fused,and a new method for scoring the hypothesis of instance segmentation is proposed,which is very important for the evaluation of instance segmentation.In order to form an effective feedback between the mask generation network and the mask quality evaluation of the whole image instance segmentation network,which is conducive to the back propagation of the mask generation network,the mask quality score and the mask loss function are combined to monitor and evaluate the generated mask quality,and an effective positive feedback is formed between them,so that the generated mask quality is within the scope of the mask quality supervision network It is more refined under the action of the.