Research On Optimal Algorithms For Object Detection Based On Deep Learning

Object detection is a fundamental problem in computer vision and robot vision system,the researches of which have obtained a significant progress in various applications.Like many other computer vision problems,there's still much room for improvement.This paper basically emphasizes on top-down attention-based optimization and adaptive non-maximum suppression to enhance the detection performance.The specifics as follow:Attention map-based optimization of object detection is a more effective object detection algorithm as well as a algorithm closer to the biological vision mechanism.Focus on the constrain on performance of object detection where the mainstream single feed-forward object detection pipeline cannot combine the high-level semantics with low-level features,this paper augments the Faster R-CNN detection pipeline with bottom-up and top-down information to help improve the performance of object detection,the main studying contents of this article included these aspects:incorporating the top-down attention into object detection network,which generates an attention map for each input image about targets.And in specific fusion optimization,we design a new ranking parameter with attention map and foreground score,which reflects the relevancy between the bounding boxes and targets,and in this way,object-relevant bounding boxes would be selected while those of non-relevant would be suppressed,which improves the accuracy of object detection.Adaptive non-maximum suppression proposes an adaptive bounding boxes fusion formulation,which could handle the object detection task in crowded scenes effectively,as well as is a more biologically plausible fusion formulation,the main studying contents of this article included these aspects: Based on attention map,we propose adjacent object differentiation to compute the probability of missing object in overlap region,and based on the adjacent object differentiation,we design the linear and non-linear score decay function respectively,and in this way,in each ergodic for bounding boxes,each bounding box's ranking score would be decayed according to their adjacent object differentiation and overlap ratio with top score bounding box,which could well select the bounding boxes containing different object while suppress those repeated detected bounding boxes,and the proposed fusion formulation is an adaptive formulation that break the constrain of manual NMS threshold,so that the performance of object detection can be improved effectively.In experiment,we choose the PASCAL VOC and MS COCO datasets to evaluate detection performance respectively,and perform hyper-parameter analysis to evaluate our method in detail.The results demonstrate that the proposed algorithm in this paper could enhance detection performance effectively.