Research On The Saliency Object Segmentation And Classification Based On Humanoid Vision And Its Application

Object segmentation and recognition in natural scenes is an important problem in computer vision.After a long period of evolution,human vision has formed its own unique mechanism of visual information processing,which provides new ideas and inspiration for the research of computer vision.Based on the theory of visual saliency and the combination of superpixels segmentation and contour detection model,this paper studies the segmentation and classification of salient objects according to the working style of human visual system.The main contents are as follows:In chapter 1,the research background,significance and purpose of object segmentaion and classification are expounded and the research status of related theory in object segmentaion and classification are summarized.Besides,some existing problems in object segmentaion and classification are analyzed and the main research contents of this paper about those problem are given.In chapter 2,how to segment multiple saliency objects in natural scene is studied.Based on the attention-shift mechanism of human vision attention model,saliency objects are detected in a natural scene and segmented through graph cut model.The main steps are as following.Firstly,the boundary of an image is detected and represented in a hierarchical region tree.All the leaves of the hierarchical region tree work as proto-objects.Then the proto-object,which includes the most saliency point in the detected saliency map,is chosen as the foreground seed for the current saliency object.Finally,the detected saliency object is segmentd to various regions by graph cut algorithm with various parameters and features.Through the application of attention-shift mechanism,scenes with multiple objects could be successfully handled.Experiments show that good results could be produced by the proposed model for natural scenes with multiple targets.In chapter 3,in order to improve the efficiency of the processes of traditional boundary detection methods,new superpixels based boundary detection methods are introuduced and applied to saliency object segmentation task.Firstly,superpixels are detected for a given image and the boundaries of superpixels are dilated to compensate for the possible localization error between the superpixels boundaries and the saliency object boundaries.Then the local boundary information of different directions and scales is extracted from the dilated superpixels boundaries and is linear weighted to get the local saliency object boundary detection results.Besides,based on the local boundary information,a symmetric similar matrix is built and resolved by generalized eigenvalue decomposition theory to achieve its feature vecotrs.Then,gobal boundary information is obtained by using gradient operator in different directions to those feature vectors.Finally,the global boundary of the object is obtained by linear weighting the local and global boundary information.Because only patially image pixels are considered to take part in the process of boundary detection,it reduces redundant computation and improves the efficiency of boundary detection.Experiments show that the proposed methods not only have a higher efficiency than the original method,but also preserve a competitive quality in the application of both boundary detection and saliency object segmentation.In chapter 4,based on the original graph-based superpixels detection algorithm,a more efficient superpixels detection algorithm is proposed and the performance of different superpixels algorithms applied to saliency object boundary detection is also studied.Instead of pixels used in traditional graph-based method,regions produced through watersheds are used to construct graph in the proposed method.The experimental results demonstrate that although the proposed method includes an additional process of watershed detection,the final efficiency is still superior to the traditional graph-based method.Besides,the performance of different superpixels detetcion algorithms applied to contour extraction is studied with the hope to find a more appropriate superpixels detection method for saliency object segmentation.Experiments are started in two different aspects.On the one hand,the boundary of superpixels itself is taken as the finally saliency object boundary extraction result.In this case,experimental results demonstrate that the performance of superpixels with irregular shapes and different sizes is significantly better than superpixels with regular shapes and similar sizes.On the other hand,various superpixels detection algorithms are applied to a local saliency boundary detection algorithm.Experimental results show that some of the superpixels with regular shapes and similar sizes could also obtain good contour detection results because of the consideration of gradient information.In chapter 5,in order to achieve both segmentation and category for a saliency object,object classification task is integrated in the work of saliency object segmentation.The main steps are as following.Firstly,multi-scale segmentations of a saliency object are produced by a saliency object segmentation model.Then the category of current saliency object is given by a learned multi-calss linear classifier.Specifically,for each segmentation of a saliency object,a feature descriptor is extracted from three different regions:the segmentation itself,its corresponding superpixels and the whole image.Based on the descriptor,the category of current segmentation is given by the learned classifier.Since every salience object has multiple segmentation results,multiple classification results will be given for each salience object and final category of current saliency objects will be determined by voting.Finally,based on the category of the saliency object,the segmentation result with the highest score is selected from the multi-scale segmentations using the linear regression model of the corresponding category as the final segmentation for current saliency object.Experimental results show that the proposed model could achieve good performance in a natural scene.In chapter 6,a set of saliency object segmentation and classification system is built according to the research contents of this paper.Five functional modules are realized in our system.These modules are responsible for saliency map extraction,visual attention shift,saliency object boundary extraction,saliency object multi-scale segmentation and saliency object segmentation and classification.The proposed system is tested using images acquired by an industrial camera.In chapter 7,a brief summary on this paper is made.This paper briefly discusses the main research contents,conclusions and innovation of our research,and makes a prospect for the next research work.