Objectness Proposal Based Semantic Object Detection And Seg Mentation

With the development of intelligent device and social media,the image data on Internet has grown with rapid speed.Therefore,the automatic analysis and processing of these data through computers poses an interesting but critical challenge.Among them,object detection and semantic segmentation are two of the most basic research issues in the field of computer vision,and scholars and experts have also conducted in-depth studies and extensive field researches for decades.Object detection technology aims to detect target objects with the theory and method of image processing and pattern recognition,to determine the semantic category of these objects,and to mark the specific position of objects in the image.Semantic segmentation describes the process of associating each pixel of an image with a category label.Semantic segmentation is the understanding of images at the pixel level,where we want to assign an object class to each pixel in the image.In recent years,Convolutional Neural Networks(CNN)have shown remarkable performance in many computer vision tasks.They have achieved a series of major breakthroughs,and have even exceeded human-level performance in some visual tasks.Most of these CNN-based object detection methods are based on object proposals,which are generated through low-level visual features efficiently,and then are fed into CNN for semantic classification.However,the object proposal methods are lacking location accuracy,as well as with large redundancy.In terms of semantic segmentation,most current CNN-based methods require pixel-level annotation for supervision,however,such annotation is very time-consuming and labor-intensive.To address them,this paper makes the following contributions:1.A new geodesic saliency detection in contour presented to improve the quality of object proposals.Based on the obtained saliency map of each candidate box,which get a refined box with better localization than the initial one.By further applying multi-sizes in saliency detection,the input candidate is well refined with both high diversity and accurate localization.Finally,boxes with high objectness are pop out by their saliency scores.By integrating our saliency refinement,all the existing methods are improved by a large margin in PASCAL VOC 2007 test dataset both at high IoU threshold and few proposal number,which can benefit subsequent object detection task.2.Making a comparison of the existing unsupervised objectness measuring methods to evaluate their effectiveness and generalization abilities.By extensi-ve experiments on popular PASCAL VOC 2007 dataset,making a comparative study of these proposal re-ranking methods and try to answer which one can achieve best accuracy in measuring objectness for each proposal.With the help of an accurate proposal re-ranking method,Non-maximal suppression(NMS)can be further applied to reduce proposal redundancy,and finally greatly benefit subsequent object detection task.Thus,research results could have a significant impact for applications required high quality proposals and high run speed.3.A simple semantic segmentation method using region-based object detector,in which only bounding box annotations are used in the training of object detector.Thus,it can be applied into real application more easily and more training data can be used to further improve performance.Experimental results in PASCAL VOC 2012 validation dataset show its comparable performance with fully supervised methods.Which can get more accurate segmentation results near the object edges which is contributed by the accurate contour detection.