Multi-scale Object Detection Network Based On Weak Supervised Deep Learning

With the wide application of artificial intelligence technology in the field of computer vision,object detection as one of the representative problems in computer vision has been paid more and more attention.Object detection technology from the most primitive use of traditional image features and machine learning methods,and gradually developed to today's use of deep learning methods,and the effect of detection is gradually improved.In the problem of object detection,the multi-scale characteristics of the object to be detected are often a difficult point in the detection process,aiming at this problem,this paper mainly studies a multi-scale object detection network based on the deep learning of weak supervision,and designs and implements a real-time object detection system based on this.Taking the cultural relics in the museum as an example,this paper has carried out a series of studies on the problems encountered in the research process of this subject,the main work is as follows:(1)An improved YOLO v3 multi-scale object detection algorithm is proposed.By using the strategy of deformable convolution and enhancing the receptive field on the basis of the original algorithm,the detection effect of the original network on the multi-scale object is improved.In addition,the use of Group Normalization strategy in the network alleviates the problem that the batch size is too small due to Batch Normalization in the network,which affects the performance of the network.and by using the label smoothing regularization method to further improve the classification performance of the original network.In the process of input data processing,histogram equalization technology is used to improve image contrast,which improves the detection effect of the network on those images with low foreground and background contrast.(2)A object detection network based on weak supervised deep learning is proposed.The basic principle is to use a specially designed dual-branch network.The network can perform the training of the classification and positioning tasks only by using the classification loss,so that the learning of the object detection task can be completed by using the image category information.And unlike other weak supervisory detection networks,the network achieves true end-to-end training by pre-setting object candidate regions without separately performing candidate region extraction operations on images before training.(3)A object detection system based on weak supervision depth learning is designed and implemented.The system is mainly composed of video streaming module,image processing module,object detection module and display module.The core of the object detection module is a multi-scale object detection structure based on weak supervised deep learning,which uses the target detection network based on weak supervision to complete the sample labeling task,so as to reduce the cost of the whole system.In addition,the part of the structure that performs target detection uses an improved YOLO v3 multi-scale object detection network.By matching the object detection module with other modules,a real-time object detection system with low cost and good detection performance is finally realized.Finally,the practicability and efficiency are proved by applying the system on the cultural relics data set.