Salient Object Detection Method And Application Research

Visual saliency model aims at mimicking the mechanism of human visual attention to perceive scene,which is widely used in many visual tasks,such as image segmentation,image retrieval,object recognition,object relocation and event recognition.Meanwhile,driven by the application of computer vision,some visual saliency models are designed to identify more informative or attractive object regions in images or videos for further processing.Therefore,how to quickly and effectively acquire meaningful objects from the image or video,and suppress or ignore uninteresting backgraound regions,that is,salient object detection,has become a core technology and preprocessing step in many visual tasks and applications.Conventional image saliency detection algorithm based on bottom-up mainly exploits various heuristic prior clues(such as contrast,background,objectiveness,center and compactness,etc)of low-level visual features(color,brightness and texture,etc)to highlight the saliency object in the scene.However,in the face of complex scenes,such as cluttered background or low contrast(objects and background regions have similar visual appearance),due to the lack of high-level semantic understanding of the scene background or objectness information,it is difficult for the saliency detection algorithm based on low-level visual features to accurately locate the complete saliency object.Compared with static images,video provides rich sequence information and continuous motion clues,which can better describe the characteristics of the object in a dynamic way.Moreover,compared with static visual features,motion cues have great advantages in dealing with clutter or complex scenes.Therefore,the conventional video saliency detection algorithm mainly employs motion segmentation to assist or fuse spatial(image)saliency to discern saliency objects in the scene.Nevertheless,due to the complicated or intermittent movement of non-rigid objects,as well as the interference of additional dynamic background and static objects,it is a very challenging task to extract salient objects that fully satisfy the constraints of temporal and spatial consistency.In view of the above-mentioned problems of saliency detection,this dissertation respectively contructs in-depth analysis and research on the saliency object detection algorithms based on image and video,and proposes novel and effective saliency object detection algorithm for image and video.Moreover,the application of image saliency detection algorithm in intelligent transportation is deeply studied,and an effective algorithm for nighttime vehicle headlights recognition based on saliency detection guidance is proposed.The main research contents and contributions of this dissertation are introduced as follows:1.An algrithm for salient object detection based on objectness guided and enhanced is proposed.To overcome the problems of traditional image saliency detection,such as background clutter or complex scenes,and the foreground object and background have similar visual apperance,it is difficult to obtain accurate and complete saliency object,this dissertation proposes to utilize high-level visual features to guide image saliency detection to improve the overall performance of saliency detection.Object proposals,as objectness information containing a series of candidate object regions,can provide high-level semantic understanding of complex scenes,which retain the integrity of foreground targets to the greatest extent,and suppress the background interference region.However,each image has different number of target suggestions,and the segmentation quality varies greatly.Directly integrating the obtained object proposals may introduce too many background noises,especially when the object proposals method cannot predict the real object regions.Therefore,this paper proposes a bottom-up salient object detection algorithm that integrates multiple saliency prior clues and the objectness information.First,the proposed mutual information-based regional consistency method is exploited to fuse multiple saliency prior clues,such as background prior,regional contrast and compactness assumption,to acquire the initial saliency map.Then,considering that many object proposals can not cover a real object accurately,an optimization algorithm based on convex function is proposed to derive(select)the optimal candidate object proposals.Three factors are considered in the optimization selection of candidate object proposals:(1)Saliency value of object proposal in image;(2)Appearance similarity between object proposals;(3)Spatial distance constraint between object proposals.Meanwhile,the optimized selection parameters based on the sparse constraint are employed to fuse multiple candidate object proposals to generate the corresponding objectness map.Finally,a saliency optimization framework under the guidance of objectiveness information is exploited to fuse the objective map and the initial saliency map to generate the final saliency map,considering the spatial smoothness constraint.Experimental results on different data sets show that the proposed algorithm based on objectness-guided image saliency can better suppress the interference of background while highlighting the completeness of saliency objects and the consistency of estimated saliency value.2.A salient object detection algorithm based on moving object properties is proposed.To overcome the problems of traditional video saliency detection,such as complex scene or cluttered background,and the intermittent motion and partial motion of non rigid objects,it is difficult to acquire complete and spacetime-consistent saliency object,this dissertation proposes to fully utilize the complementarity between the instance-level object segmentation and the salient motion estimation derived from the sequence images to locate the motion objects,while suppressing the interference of dynamic background and static objects.First,the object instance and motion vector estimation of the current frame are obtained by using the pretrained object instance segmentation model YOLACT++ and the optical flow estimation model Flow Net2.0 respectively.Meanwhile,based on the obtained motion vector map,the saliency detection method is exploited to generate the moving detection results of the current frame.However,both the results of object instance segmentation and salient motion detection may have low confidence at the same time.In order to obtain accurate and effective moving object regions,this dissertation proposes to utlize the strong visual appearance correlation and consistency of the object(background)region between neighboring frames and the whole video sequence to correct and improve the results of object instance segmentation and salient motion detection.The specific steps are as follows:(1)Based on the constructed adjacency matrix,using different key frame selection strategies and the nearest neighbor priority principle,the high-quality detection results of the selected key frames are propagated to remaining non key frames in an orderly and iterator manner through inter-frame sparse reconstruction,which improves the overall quality of initial object instance segmentation and salient motion detection,while avoiding the problem of excessive accumulation of errors caused by the traditional inter-frame consistent propagation algorithm.(2)By using the integration strategy based on multiplication,the improved object instance segmentation and salient motion detection results are fused to extract the image regions with both object attributes and moving attributes to suppress the interference of dynamic background and static object.(3)Based on the obtained initial information of the moving object,the bayesian fusion framework incorporating the global features of moving object and background region that are obtained through the low-rank constraint is utilized to further improve the global consistency of the moving object estimation.Then,in order to obtain more prominent and complete saliency objects in space,a method based on geodesic distance transformation is employed to obtain the spatial saliency estimation of the current frame based on the prior information of moving objects and background regions.Finally,in order to obtain spatiotemporal consistent saliency objects,a graph based energy optimization function is proposed,which effectively integrates multiple saliency clues,such as moving object clue(temporal object information),spatial saliency clue(spatial object information)and global features,to complete the optimal assignment of saliency estimation.Quantitative and qualitative experimental results on different datasets show that the proposed algorithm can suppress the interference of similar background region and static objects in the case of clutter scenes and complicated motion while avoiding excessive diffusion of error detection results.3.A nighttime vehicle headlights detection algorithm based on saliency detection and PHOG features is proposed.In the nighttime traffic scene,the distance between the vehicle and the camera is different,which results in the different size and brightness of the vehicle headlights in the traffic images.Meanwhile,the interference of various reflected light(such as ground-surface,water-surface,vehicle-bodies or lane-marking)make it difficult for traditional methods based on template or thresholding to detect faint and small-size headlights while maintaining the complete shape of the headlights.To overcome the above mentioned problems,a multi-scale local saliency detection method based on maximum value is proposed to extract candidate headlight bright blocks.Then,in order to improve the recognition ability of vehicle headlights and make full use of the structure information of headlight,the gradient map of the candidate headlight block is encoded by space pyramid representation to construct the PHOG(Pyramid histogram of oriented gradients)feature,which is insensitive to geometric and optical transformations.Then,the pretrained SVM classifier is used to effectively classify the candidate headlight blocks based on the PHOG features,and the vehicle headlights are accurately discerned.Finally,through the quantitative and qualitative comparison of different vehicle headlights detection methods in different scenes,the results show that the proposed vehicle headlights detection method can effectively improve the accuracy of headlights detection,and suppress the interference of all kinds of reflected light.