An Integrated Approach For Joint Visual Object Detection And Sub-category Recognition

Visual object detection refers to the automatic object localization in videos or images by computers and sub-category recognition refers to the further classification of the above objects.Both visual object detection and recognition are long-term focuses in the field of computer vision.It has a great significance to accomplish fast and accurate object detection and recognition.Object detection and sub-category recognition are usually separated into two independent but sequential processes in most researches,in which recognition follows after detection.As a matter of fact,object detection and sub-category recognition have a certain degree of information coupling instead of being separated completely.Some information in the process of the object detection can help sub-category recognition.In return,some information in the process of recognition may help to confirm or reject the results of detection.But the traditional process does not make use of these informations.Therefore,it is of great significance to do some researches on how to fully utilize those informations and complete the two processes better and simultaneously.As an object detection approach,deformable part model(DPM)has attracted a lot of attentions for its creative design and excellent performance in the field of object detection.DPM can provide many local details and sub-category information of objects by introducing the concepts of part filters and mixture model,which creates possibilities for the integration of the object detection and sub-category recognition.However,high computation time cost of DPM brings difficulties into real applications.Taking all of the above into consideration,this paper mainly completed the following tasks.(1)The process of feature extraction in DPM was accelerated using graphic process units(GPU).It's about ten times faster than the CPU version.(2)An integrated method based on score bias for object detection and sub-category recognition is presented.Score bias was introduced to balance scores from different models allowing simultaneous detection and recognition.And a software framework for vehicle information detection and recognition was built.(3)An integrated method based on learning for object detection and sub-category recognition is presented.A multi-objective optimization problem was introduced using structural support vector machine with latent variables.The overall solution to the new problem was given by strict mathematical analysis.On the self-built vehicle dataset,both the precision and recall of detection and accuracy of recognition can reach higher than 95%.