Researches On Visual Saliency Detection Via Multi-level Image Feature Integration

In recent years,the rapid development of mobile Internet and social network has resulted in an explosion of digital images.The interesting regions can be selected by computer employing salient detection technology.Therefore,salient detection,which make image processing more efficient,become research hotspot in computer vision.However,the existing salient detection algorithms are still insufficient on the salient object integrity and background noise suppression.Based on visual salient detection principle and feature integration theory,in this paper the novel salient detection methods are proposed via multi-level feature extraction and integration.The main work of this paper contains following points:(1)Most of existing algorithms measure saliency utilizing global features with single level.Due to the lack of local information,global features can highlight the saliency of objects uniformly,which however cannot suppress the background noise and highlight saliency of objects accurately.Therefore,based on multi-level feature extraction,in this paper the salient detection algorithm is proposed via multi-level image feature integration.Firstly,the global level feature and local level feature are mapped to the surrounding level by the designed surrounding aggregation method,and the surrounding information is complemented by surrounding level features.Then,the image is decomposed into local,surrounding and global level features,which solve the problem of information loss caused by the single level feature.Finally,the use of support vector machine for feature integration avoid the problem of parameter selection when using the mathematical operation for feature integration,and support vector machine can automatically allocate parameters through learning.Experimental results on three widely used public benchmark datasets show that the mean absolute error of the proposed algorithm perform much better than other classical algorithms,and the other performance is also better than the classical algorithms.(2)The salient detection algorithm via multi-level image feature integration has a satisfactory effect in salient object detection,but there are still some improvements in the integrity of salient object detection and suppression of background noise due to the insufficient granularity of the multi-level features.Therefore,region level and center-edge level features are complemented,and the image is decomposed into five level features.The multi-level details of the superpixel is further preserved by the region level features.Then,center-edge level features is used for noise suppression in background and edge where the salient object is more prone to the image center than the edge.In addition,the weight of the superpixel closer to the contrast center is higher in global color contrast,but the region in the salient object center has smaller contrast with surroundings and in this situation global color contrast performs not well.Therefore,a new color contrast feature is proposed by giving the region far away from contrast center more weight and it can highlight the entire salient object more complete.Experimental results show that the performance of the proposed algorithm on the mean absolute error and F-measure outstand with previous proposed algorithm and other classical algorithms.