Research On Image Cosegmentation Based On Convolutional Neural Network

Image segmentation is one of the key tasks in the research field of computer vision,and it is also the basis for image recognition,image analysis,and image understanding.In recent years,especially after the introduction of deep learning,researchers have proposed a variety of effective image segmentation schemes,which can ensure better segmentation performance in the presence of a large number of high-precision pixel-level image labels.However,obtaining such labels requires high labor costs,so,in practice,there are often image segmentation tasks with missing accurate labels.To this end,researchers turn to weakly supervised image segmentation and use weak labels to directly implement object region segmentation.Because the weak labels are easy to obtain and the labeling cost is low,the research and application fields of weakly supervised image segmentation methods are more wide.However,the information provided by weak labels or image-level labels is relatively rough,so it faces the challenge of how to implement pixel-level segmentation under rough labels.The image co-segmentation method is the core content of weakly supervised image segmentation.It aims to segment images using image-level labels and common information between images.Although artificially designed features and deep convolutional features are applied to image co-segmentation,the performance of segmentation is still insufficient,because the prior information provided by image-level labels is relatively rough,and it is difficult to accurately locate the common object area.For this reason,this paper introduces a small amount of pixel-level labels,which significantly compensates for the lack of information provided by weak label without increasing the burden of manual labeling.At the same time,a more robust image co-segmentation model was constructed using convolutional neural networks.The specific research contents are as follows:1.For the existing problem of insufficient priors only considering image-level labels,this paper proposes a two-branch image co-segmentation network based on a small amount of guidance.The network not only introduces a small number of pixel-level labels to provide accurate prior information,but also constructs a similarity loss function to combine two branches to segment the common areas of the image.2.For the problem that the above work only uses the limited common information provided by a single pair of images as guidance,this paper proposes a three-branch image co-segmentation network based on multiple pairs of collaboration and a small amount of guidance.Two branches of the network introduce a small number of pixel-level labels to guide the other branch to segment the image.The formation of three co-segmentation pairs by pairs of branches can not only make full use of the limited pixel-level priors,but also improve the learning rate and segmentation performance of the network through the joint training of multiple similarity loss and cross entropy loss functions.3.For the current problem that image co-segmentation is limited to segmenting images of specific categories,this paper proposes a small sample co-segmentation network based on the reverse guidance network and the attention module,using the idea of image co-segmentation to build a reverse guidance network,and putting forward the attention module to focus on common areas,so that the segmentation model can be generalized to a new category using only one pixel-level label after full training.In this paper,three datasets,such as iCoseg,MSRC and PASCAL VOC 2012 have been verified,and their excellent performance proves the effectiveness of our algorithms.