Font Size: a A A

Research On Object Detection Method Based On Convolutional Neural Network In Complex Scenes

Posted on:2020-11-01Degree:MasterType:Thesis
Country:ChinaCandidate:X N WangFull Text:PDF
GTID:2518306500483254Subject:Computer Science and Technology
Abstract/Summary:PDF Full Text Request
The existing common object detection algorithms only have better detection performance for general scenes,but are not applicable to complex scenes.Object detection in complex scenes has the following problems: 1)there is noise in the background affecting the accuracy of object detection;2)the detection object has many complex components;3)there is only a very small difference between the object and the background,and the colors and other features are very similar.These problems pose great challenges for object detection.How to achieve object recognition and detection from complex scenes becomes a more important and difficult problem.This paper focuses on the object detection algorithm in complex scenes,and proposes the Object Detection in Complex Scene(ODCS)based on convolutional neural network to reduce the detection error of the object detection algorithm.The main innovations of ODCS are as follows: 1)A RoI feature selection network is proposed.By using the difference between the original region and the aspect ratio of the original RoI to change the feature representation of the RoI,a sub-region attention bank and an aspect ratio attention bank are generated for the entire image.The RoI-based sub-region attention map and aspect ratio attention map are selectively pooled from the bank and then used to refine the original RoI features of the RoI classification.The RoI feature selection network is based on the popular Conv Net backbone(Res Net-101,Goog Le Net and VGG-16)and is equipped with a lightweight detection subnet to continuously improve detection performance.2)By exploring rich context information to improve the detection model,a context inference network is designed to iteratively propagate information between different objects and the entire scene.The context inference network can consider not only the visual appearance of an object,but also two contexts,including scene context information and object relationships in a single image.When using these structured information,object detection is considered a cognitive problem and a reasoning problem.This paper examines the performance of the ODCS algorithm through two sets of data sets.Firstly,the performance of Faster R-CNN,YOLO,SSD,R-FCN and ODCS on the PASCAL VOC dataset was compared.In addition to the comparison of detection accuracy,another comparison was provided to evaluate their test consumption on PASCAL VOC 2007.Secondly,the performance of Faster R-CNN,SSD and ODCS on complex scene datasets,including accuracy,false detection rate,number of iterations and error rate,is compared.Both experimental results show that the proposed ODCS algorithm can better solve the interference problem caused by complex scenes in object detection,and the object detection accuracy is obviously improved.The algorithm effectively improves the detection accuracy of objects in complex scenes,especially on test images with complex backgrounds.
Keywords/Search Tags:Complex scenes, Object detection, Convolutional neural network, RoI feature, Contextual information
PDF Full Text Request
Related items