Research And Implementation On Co-saliency Detection Algorithm Based On Deep Multi-Instance Learning

Co-saliency detection is an emerging and rapidly developing research direction in the field of computer vision.As a branch of saliency detection,co-saliency detection refers to obtaining foreground regions from two or more related images.And these foreground regions are both salient and consistent.Also,co-saliency detection has been well applied in other fields.At present,there are two kinds of co-saliency detection methods.The first is the traditional co-saliency detection method,which relies heavily on the prior knowledge of human beings,and calculates the co-saliency of image groups based on the characteristics of artificial design.And the method is difficult to adapt flexibly to the complex environment in the real application scenarios.The other methods based on the unsupervised learning,which are pursued by most researchers.Although unsupervised learning is well applied in many directions,the models designed to study the problem of co-saliency are often ignored the weakly labeled message of each image.As a result,the trained model lacks robustness and does not show good results in actual scene applications.Based on the analysis and research of the above two problems,this paper proposes to solve these problems of co-saliency detection using multiple instance learning.In order to build a complete detection framework,this paper proposes to solve these problems of co-saliency detection using deep multiple instance learning.First of all,in order to solve the dependence on handcraft features,this paper uses deep neural network to extract the features of the image to reduce the influence of human knowledge on the performance of the algorithm.Then,in order to make better use of the weak label information carried by the image,this paper uses multi-instance learning to solve the problem of co-saliency detection,which is the classic method of weakly supervised learning.Finally,this paper gets co-salient objects of image groups by training a deep multi-instance learning network.Specifically,the main research work of this article is divided into the following two parts:(1)We transform the problem of co-saliency detection of image groups into a multi-instance learning problem.Aiming at the current problem of insufficient utilization of weak label information of the dataset in unsupervised learning,the co-saliency detection is transformed into multi-instance learning.The weak label information carried by the image group is used as the basis for constructing bags and instances.And the multi-instance learning algorithm is used to assign labels to the instances without label information in the package.And then the positive instances containing the object of co-saliency are selected out.Finally,we obtain the co-saliency object of the image group.(2)Co-saliency detection is performed using a deep multi-instance learning network.This paper combines multi-instance learning with deep learning to train a deep multi-instance classification network.This paper uses convolutional layers in the network to complete the extraction of bags and instances features,reducing the interference of human factors on the performance of the algorithm.We use group-wise semantics as top-down semantic guidance to promote co-saliency bottom-up reasoning.The method based on instance clustering is used to optimize the deep multi-instance network to improve the performance of the model.Finally,this paper conducts experiments on iCoseg and MSRC,which are commonly used for co-saliency detection,and calculates the evaluation indicators to verify the performance of the algorithm.