Visual Salient Object Detection Based On Structural Constraints And Its Applications

Extracting salient information from online data has become a popular research area in computer vision since the boost of internet technology and massive growth of multimedia data.The human visual system could automatically select the most important area in the scene,and only process the selected area while ignoring the information from other areas.Inspired by this,the researchers proposed a visual saliency computation model to simulate the human visual attention mechanism.Salient object detection as a significant research branch of visual saliency computation model,aiming to detect the most important object in a scence,has received more and more attention and has made considerable progress in the past decades.It has been widely applied in various computer vision tasks such as object segmentation,object tracking,and image compression.This dissertation focuses on the salient object detection methods and its applications,the spatial relationship of different superpixels,multi-scale and other structures in the image are constrained,proposes two salient object detection methods and an attention mechanism-based lane line detection model,and the saliency detection technology is applied to the road available area detection task.The main work of this dissertation could be summarized as follows:(1)A salient object detection method based on background-absorbing Markov chain is proposed.The proposed method first removes the most different edge from the four boundaries of the input image,and the nodes on the remaining three boundaries are selected as the absorption nodes of the Markov chain.After calculating the absorbing time from the transient nodes to the absorption nodes,the initial saliency map is obtained.Then,the probably background nodes are selected from the initial saliency map by a threshold,which are duplicated as absorbing nodes,and calculate the absorption time to get the saliency map.Finally,a multi-layer of saliency maps are used to achieve the eventual saliency map.The proposed method has been proved to optimize the accuracy of detection and achieve the more accurate saliency maps and it is more suitable for images where salient objects occupy the boundaries of image.(2)A salient object detection method based on bidirectional absorption Markov chains is proposed.The input image is first segmented into a number of superpixels,and the four boundary nodes(duplicated as virtual nodes)are selected.Subsequently,the absorption time upon the transition node's random walk to the absorbing state is calculated to obtain the foreground possibility.Simultaneously,foreground prior(as the virtual absorbing nodes)is used to calculate the absorption time and get the back-ground possibility.Besides,the two aforementioned results are used to form a com-bined saliency map which is further optimized by using a cost function.Finally,the superpixel-level saliency results are optimized by a regularized random walks ranking model at multi-scale.The comparative experimental results reveal the superior performance of the proposed method which could overcome the single directional Markov chain method and could highlight the region of the salient object.(3)A road available area detection method based on fusing the attention area and saliency area is proposed.The proposed method first obtains a candidate vanishing point in the image and creates an attention area in the candidate node's round in the driving scene.Then,the saliency detection algorithms are used to obtain a salient area in the scene,and the final vanishing point is obtained by fusion of the attention area and the salient area according to the prior information in the road scene.Subsequently,the position of the final vanishing point and the two boundary nodes of the road edge are determined to build a triangular road area.Finally,the saliency detection algorithm is used to calculate the saliency area in the triangular road area to obtain the saliency area,and then the saliency area is removed from the triangular road area to get the road available area.The experimental results show that the proposed method can overcome the shortcomings of the traditional vanishing point detection algorithm,reduce the complexity of the algorithm by using the attention area and the saliency area.(4)A multi-class lane lines detection method based on the attention mechanism with multi-scale structures is proposed.The proposed method is based on Deeplabv3+network architecture and has encoder and decoder structures.Firstly,the atrous convolution at multi-scale by applying attention mechanism is designed in the encoder module to obtain more features.In the decoder module,the semantic embedding branch module is used to combine the high-level and low-level semantic information to obtain more rich features and the single stage headless context detection module is used to obtain the multi-scale lane lines features.Experimental results show that the proposed method can be applied to lane line detection in driving scenes,and could be used for semantic segmentation of multi-class lane lines,and has a good detection result.