Research On Algorithms For Fine-grained Image Segmentation Of Scene Understanding

With the great development of Artificial Intelligence,like the invention of series of original products-unmanned automobile,the requirement of analyzing-image ability has gradually increased.Image segmentation is playing an important role in many different fields.The result of image segmentation is helpful for subsequent scene understanding and analysis.Besides,related research has a vital meaning and comprehensive applied scenes.Image segmentation is to distinguish each category of objects by calculating masks for each object in the image.The fine-grained image segmentation algorithm doesn't only calculate masks for each object instance in the image,but also needs to distinguish classification information of each object instance in the image.It requires the algorithm to fulfill the segmentation during the fine-grained classification,which is helpful for subsequent scene understanding and analysis,such as a scene-analysis attention end-to-end module.Compared to traditional image segmentation methods,the related study shows a bigger challenge.The thesis is mainly about three tasks mentioned below.(1)The study analyzes and compares four algorithms which are based on deep learning and get great results of the segmentation.They are FCN,SegNet,FCIS and Mask R-CNN.Among them,FCN is the first one to apply fully convolutional network structure to semantic segmentation tasks,which is a method of segmentation end-to-end and pixel-to-pixel;SegNet network is similar to FCN network,but they are different in the network of encoding and decoding;FCIS presents a feature-extraction method based on two translation variants,object features and background features.Object features are used for segmentation and background features are used for classification;Mask R-CNN completes two tasks in the same network:object detection and instance segmentation.The study shows Mask R-CNN has a more promising result of image segmentation,which is the base of the backbone network module.(2)We come up with a Feature Pyramid Attention mechanism(FPA).From chapter 2,we know that all present methods of image segmentation have a poor learning ability of pixel location information.Through pyramid structure,the proposed algorithm,FPA,pays attention to Mask,learns from Mask instead of feature maps and then preserves information of pixel location to the maximum extent,which helps us to make a complete use of context information and furthermore,improves the quality of generating Mask.The study shows the algorithm of the Feature Pyramid Attention mechanism has a better result of small-scale image segmentations.(3)We present a fine-grained image segmentation algorithm based on the Global Feature Pyramid Attention(GFPA).Traditional algorithms can only distinguish instance category and generate Mask.However,they cannot distinguish fine-grained classification information of the same big category.In order to fulfill fine-grained image segmentation,we have improved the FPA module mentioned in chapter 3 in the way of adding a global pooling module to FPA,which makes up of GFPA.We adopt the database of Open Images V4 to train our branch of classification network in order to get the result of fine-grain segmentation.The database of Open Images V4 only marks image-level labels and objective bounding boxes,so during the training of the branches of Mask,we adopt the labels of Mask in the database of COCO to train our network.Through the cooperation and transferred knowledge of two networks,we can fulfill the fine-grained pixel segmentation.The study shows,compared to Mask R-CNN,the algorithm in our thesis can fulfill the fine-grained classification and the pixel segmentation at the same time.