Research On Saliency Detection Based On Feature Fusion

The goal of computer vision is to understand the content of a given image or video correctly.To this end,it is necessary to determine the location of the main object of the image,which is the research content of the image saliency direction.The research of image saliency was initially based on the simulation of human gaze using simple filtering and directional or color features.Such traditional saliency detection method exploiting manual features has achieved good results on open data sets.With the development of depth learning,especially in image recognition competitions,which gradually surpasses the human recognition result.A large number of features extracted from those depth learning models are widely used in the field of computer vision,including image saliency detection.At the same time,the current stereo vision multimedia equipment is growing,and the algorithms for stereo images are also being developed.As a basic algorithm processing step,stereo vision saliency detection has also received more attention.At present,some saliency detection algorithms for stereo vision have been proposed,but there are still many problems.For example,these saliency detection methods can not balance various features,and the detection speed is too slow.In addition,the research on the use of depthinformation is still in its infancy.To solve these problems,this paper attempts to use a more comprehensive and unified method to detect saliency.In the design of the algorithm,the depth information and color information are processed in a holistic manner to make full use of the correlation between each channel.In addition,several end-to-end stereo vision depth learning models are established.These models use the weights from image classification depth learning model as the pre-training weights to ensure that features for saliency detection task can be sufficient.The four models presented in this paper are tested in multiple stereo visual salience open data sets.The test results show that the proposed two-channel depth learning model and the coded decoding network model exceed the existing algorithm by more than 0.1 on CC and KLDiv metric.Other indicators are also superior to existing methods.The proposed method processed in quaternion frequency domain with manual feature exceeds the existing method with manual feature in each evaluation metric.The parameter sharing depth learning model is superior to the existing algorithm by 0.1 in the KLDiv metric,and is also superior to the existing algorithm in the CC and KLDiv metric.It is only 0.055 less for the most suitable method in the AUC evaluation metric in the NUS database.In the NCTU database,the metric evaluation gap between most suitable model is within 0.003.