Salient Object Detection Based On Multi-level Contextal Information Extraction

The salient object detection aims to utilize the attention mechanism of Human Visual System(HVS)in an algorithm to quickly detect the more prominent object areas in an image.It has been widely used as a pre-processing step in various fields of image processing,such as image segmentation,image fusion,image cropping among other.In recent years,significant salient object detection algorithms based on deep learning have achieved good detection results in most natural scenes.However,for some complex scenarios,such as irregular targets in the scene,the existing salient object detection algorithms cannot realize a complete and consistent detection of saliency object region due to the lack of contextual feature information,.Therefore,constructing a significant salient object detection algorithm model with good detection performance is still a key issue in the field related to computer vision.On the basis of summarizing and analyzing the related research achievements in salient object detection field,this thesis has proposed a new salient object detection method based on the multi-level context feature extraction and fusion.At the begining,the thesis summarizes exsting research results related to the field of image salient object detection.Especially,three related algorithms are emphatically introduced in detail.These algorithms contains the salient object detection employing local tree-structured low-rank representation and foreground consistency,the deeply supervised salient object detection with short connections,and a stage-wise refinement model for detecting salient objects in images.Additionally,for the existing algorithms utilizing the single-scale convoltion kernels to extract multi-level feature,an issue that object detecction region missing is normally caused.Therefore,this thesis proposed a method of salient object detection based on multi-level contextual feature extraction.The proposed algorithm mainly consists of the prossess as follows.(1)The basic network is used to extract multi-level depth features from shallow to deep.(2)A multi-scale contextual feature extraction module based on deformable convolution is constructed to extract rich context features in each depth feature to ensure the completeness of target detection.(3)An attention module is constructed to weight the features extracted by the network.It can improve the effect of effective feature information,and restrainthe error feature channel from affecting the detection result.(4)A multi-level feature fusion mechanism is introduced to obtain the final feature map.(5)A boundary refinement module is used to improve the boundary of the salient object,and subsequently output the final saliency map.Finally,the algorithm is implemented in the Caffe deep learning platform under Ubuntu14.04 environment.The test program complied by MATLAB is employed to evaluate the algorithmapplied to four publicly significant salient object detection datasets.The proposed algorithm has also been compared with the existing 13 classic salient object detection algorithms.The experimental results show that the saliency map generated by developed algorithm can accurately and consistently highlight the salient objects via the comparison.Moreover,it can achieve an excellent performance in the quantitative indexes such as PR curve,F-measure curve and MAE value as well.Hence,the effectiveness of proposed algorithm in this thesis can be verified.