Research On The Techniques For Image Co-segmentation Based On Deep Learning

Co-segmentation aims at segmenting objects of a same semantic class in multiple input images,which is of great benefit to help researchers obtain objects of interest from massive network images.Co-segmentation task is also one of the current hot spots in artificial intelligence research.Although researches on Co-segmentation has been studied for a long time,there are still many problems that should be discussed based on deep learning.The current deep learning based co-segmentation models usually utilize Siamese network to extract the semantic features of a pair of images.Then feature comparison is performed across the semantic features.At last,the interacted feature should be parsed and give the locations of the common object.Although the performance of deep co-segmentation methods have been greatly improved compared to the non-deep methods,the deep methods still cannot handle well when there are multiple complex objects in the image.The complex images usually contains intra-class variations and inter-class similarities in reality.The current models usually perform low lever(pixel-wise)feature comparison and it cannot deal well with the intra-class variations and inter-class similarities.Based on the human visual mechanism,we proposes a co-segmentation algorithm based on hierarchical semantic consistency mining.The algorithm divides semantic features into multiple regions of different levels,and mines common targets based on regional feature consistency.This method has a larger feature comparison area than the previous method,and the feature covers the entire information of the target,which brings more robust feature comparison.Experiments prove that the proposed co-segmentation algorithm is of great help to intraclass variations and inter-class similarities.The previous method requires artificial design of hierarchical parameters.This paper further improves it and draws on the idea of self-attention mechanism to propose a cosegmentation algorithm based on adaptive object-level feature comparison.The algorithm can adaptively acquire the context information of whole object,which means the feature of each location can adaptively acquire the information of the entire object regardless of the size of the target area.Based on the object-level features,the model further achieves more robust and more adaptive co-segmentation.At the same time,this method further expands the dimension of feature comparison and performs common object mining on the features of different levels of the convolutional neural network.The multi-level features can provide more detailed structural information and help the co-segmentation model achieve more delicate results.Siamese network can only process a pair of images at one time.This paper further proposes a co-segmentation algorithm based on group consistency features.The algorithm draws on and improves the structure of the recurrent neural network so that it can mine the consistency information contained in a group of images.Then the group consistency information serves as a reference and guidance for each image.The group consistency information helps further enhancing the model's ability to mine common goals and achieving better co-segmentation performance.At the same time,the Siamese network needs to repeatedly select the paired image when processing a group of images.The time complexity is relatively large.Our group consistency based co-segmentation method is able to process all images at once and greatly saves time cost.