Research On One-stage Object Detection Algorithm Based On Convolutional Neural Network

Computer vision,as a discipline that studies the use of computers to obtain highlevel,abstract information from images and videos,has become one of the important research areas in the era of artificial intelligence.Object detection aims to obtain the position and size of target objects from images and videos,and it is a fundamental but challenging research hotspot in the field of computer vision,with wide applications in human-computer interaction,autonomous driving,and other fields.Convolutional neural networks are often applied in object detection as a technique for learning feature representations from data.In recent years,with the development of convolutional neural network technology,the rise of open-source datasets and the increase in computing power of hardware devices,object detection algorithms based on convolutional neural networks have been developed in a breakthrough manner.This thesis focuses on the network structure of the one-stage object detection algorithm and the regression method of the bounding box,aiming to improve the detection accuracy of the algorithm and accomplish the following research results.The SC(Split-and-Combine)module is designed to facilitate feature fusion,which splits the input feature map into several parts according to the channel dimension,then then each section is transferred to sampling branches of different depths for processing,and finally fused one by one.The SC module directs the network to more important feature information,which contributes to optimizing computational efficiency and improving network performance.The SC89 network was built based on the SC module,and the training was completed on the Image Net classification dataset.The Top-1 and Top-5 accuracies reached 77.459% and 93.597%,respectively.And the SC-FPN(Splitand-Combine Feature Pyramid Networks)that facilitates the fusion of multi-size feature information is designed to promote the accuracy of the detection algorithm for small-size objects.A novel method of adding positive samples,ASI(Adaptive Sample Increase),is designed to speed up the convergence of the network and shorten the training period.This method calculates the candidate region based on the size and center coordinates of the true bounding boxes,increases additional bounding boxes whose center points fall into this candidate region and have the same dimensions as the true bounding boxes,and makes the intersection ratio between the additional bounding boxes and the true bounding boxes above a threshold value.A data augmentation method,Admix,is proposed to address the problem of large memory consumption for network training.This method improves the performance of the Batch Normalization algorithm by fusing information from multiple images on a single image.The detection network SCDet was built based on SC89,SC-FPN,ASI,and Admix,and the three models SCDets,SCDetm,and SCDetl were obtained by scaling the network depth,and the performance was evaluated on the COCO(Common Objects In Context)dataset,with the FPS of 222.2,137.0,and 100.0,and AP of 38.2%,41.6%,and 43.1%.