Research And Improvement Of Region Proposal Generation Algorithms In Object Detection

The problem of computing category agnostic bounding boxes proposals is utilized as a core component in many computer vision tasks,which are widely used in image classification,object detection,visual question answering system.Region-based Neural Networks can achieve the state-of-art results and become the mainstream method in object detection field.The region proposal generation algorithm is an efficient method to accelerate the object detection,which has greatly improved the classifier efficiency of object detection.In the case of category agnostic,region proposal generation algorithm needs to output the proposals where the object may exist,since the very little prior information,it is a very challenging task.Some methods have been extensively used in previous research.The simplest method is sliding window.The method needs to extract 6 710 ~ 10 bounding boxes.Due to the large number of bounding boxes,the calculation of feature extraction and classification of the bounding boxes area is large,which greatly affects the efficiency of object detection system.The most widely used method is based on the low-level image features,mainly Selective Search method and Edge Boxes method.Although a significant reduction in the number of bounding boxes,due to the lack of high-level image semantic information,resulting in a low average recall rate.In a recent method based on neural network,the most representative is the Sharp Mask method,which can achieve a higher average recall rate using the mask area.However,bounding boxes location optimization is not taken into consideration,and there is still a great room for improvement in bounding boxes location optimization.This work based on the framework of the region-based Convolutional Neural Networks for object detection,with a general effective region proposals generation algorithm for research purpose,concentrating on how to generate fewer and more accurate class-agnostic proposals for object detection.In order to suppress the excessive negative samples in initial windows,a combination of objectness instance segmentation and small size boxes generation is adopted to reduce the introduction of the noise windows.In order to refine the inaccurate positioning of Sharp Mask method,the position optimization module in the Attractio Net method is utilized.In the optimization process,the information of bounding boxes scores and spatial approximation is added to speed up the convergence of position optimization.In order to filter the false bounding boxes in the post-processing stage,the methods of decaying confidence score and adding prior knowledge are used to improve the hit ratio of the object detection system.The experiment mainly uses the VOC and COCO dataset.The experimental results show that this work improved method can effectively improve the average recall rate of region proposals and the accuracy and efficiency of object detection.In terms of average recall,this work attains 7~12% relative improvement in different indicators compared with the Sharp Mask method,and 1~2% relative improvement on various size indicators compared with the Attractio Net method.In the framework of Multi Path,the detection system only using the optimized 20 bounding boxes can achieve the detection accuracy of 400 bounding boxes before optimization.