Images Amodal Instance Segmentation Algorithm Based On Deep Learning

With the development of deep learning technology in computer vision,common visual recognition tasks,such as image classification,object detection and semantic segmentation,have quickly reached maturity.Instance segmentation,as an important image understanding method,aims to find all the instances in the image and classify the objects at the pixel level.It has the functions of object detection and semantic segmentation.Nevertheless,instance segmentation is still not enough to fully understand the complex surrounding environment.For example,when objects have mutual occlusion relationships,instance segmentation technology can only identify and segment the visible pixel parts,and cannot predict the occluded parts of the object.However,the human visual system is born with the ability to perceive the complete physical structure of an object,and can accurately infer all its shapes,boundaries and semantics only when the object is partially visible.Amodal instance segmentation is an extension of the instance segmentation task.Its main purpose is to predict both the visible parts and the occluded parts of each instance,so as to help the computer vision system realize the perception of the entire range of objects.This dissertation focuses on the drawbacks of current amodal instance segmentation methods based on convolutional neural networks.The main research contents are as follows:(1)The convolutional neural network performs downsampling through convolution and pooling operations with stride,which reduces the amount of network parameters,but causes serious loss of image spatial detail information,which makes the pixel-level image segmentation effect poor.To solve this problem,this dissertation proposes a bilateral network algorithm structure with spatial detail preservation and attention refinement upsampling.The method encodes multi-scale spatial detail information by designing a shallow convolutional neural network with pyramid convolution.In addition,in order to further refine the spatial detail information,the attention refinement upsampling module is constructed by introducing crosslayer information guidance.Experimental results on the COCO-amodal dataset show that the method can provide multi-scale spatial details and improve the quality of image segmentation.(2)Aiming at the segmentation underfitting problem caused by the lack of feature representation ability of current amodal instance segmentation methods and lack of global context information,this dissertation proposes a amodal instance segmentation method based on feedback attention mechanism and global context fusion.The method introduces the feedback process in the feature pyramid structure for relearning,constituting the feature pyramid structure of the feedback attention mechanism.Meanwhile,the context attention module is proposed to learn the spatial correlation and encode the global context information enables adaptive attention to more relevant areas.Experimental results on the COCO-amodal and D2 Samodal datasets show that the method proposed in this dissertation can significantly improve the prediction accuracy of the occluded parts of the object,and effectively solve the underfitting problem.