Research On Object Detection Based On Multi-level Deep Feature Fusion

With the increasing penetration of artificial intelligence into all aspects of our lives and work,computer vision,a very important branch of artificial intelligence technology,has attracted more and more attention from researchers.Object detection,as an important task in computer vision,is an indispensable prerequisite for many vision tasks,including many application scenarios such as event recognition,scene content understanding,and autonomous driving.In recent years,the field of object detection has developed rapidly,and various research results have emerged endlessly,but the current object detection technology still has a lot of room for improvement.On the base of extensive literature review and extensive research on existing research results,this thesis elaborates on the current status of domestic and foreign research on object detection and the main problems.After that,the current mainstream methods of object detection are systematically introduced,including two-stage and onestage object detection methods based on deep learning.And the main difficulties of video object detection are explained.Aiming at the problem of low accuracy in small object detection,a weighted feature pyramid network is proposed.Aiming at the problem of extremely imbalanced foreground and background samples in the one-stage object detection method,a background suppression loss function was proposed.Based on the weighted feature pyramid network,an interlaced model for video object detection is proposed.And a prototype system that integrates data processing,model training and object detection is developed.The main research contents of this thesis are as follows:1)Aiming at the problem of poor detection of small object,an image object detection method based on multi-level deep feature fusion is proposed.Using the characteristic of convolutional neural networks that shallow layers feature have higher resolution,and deep layers feature have higher specific semantic information.According to the importance of different types of information to the final object detection effect,a network structure is designed that weighted fusion of shallow layers feature and deep layers feature.The fused feature used for detection can not only make full use of the highly abstract semantic information provided by deep layers feature,but also make use of the high-resolution information of shallow layers feature to compensate for the information irreversible defect of the convolution network after repeatedly convolutions.The background suppression function effectively alleviates the problem of extremely imbalanced foreground and background samples during object detection by increasing the loss of foreground classification errors when calculating network loss.2)Aiming at the problem of low video object detection speed,a video object detection method based on interleaved WFPN is proposed.Utilizing the spatiotemporal locality of video,using high-precision feature extraction network and high-speed feature extraction network to cross-frame feature extraction for video frames,and connecting different feature extraction networks through Group ConvLSTM,so that the high-speed feature extraction network can make full use of the pre-ordered highprecision feature extraction network extracts high-precision features.In this way,it is possible to make full use of the spatiotemporal locality of the video to reduce redundant calculations,and it is possible to reasonably use the timing information of the video.Finally,a good balance of speed and accuracy on video object detection is achieved.3)A prototype system that integrates data preprocessing,model training,model pruning,and target detection is designed and developed.The system uses Python as the main development language,PyQT as the graphic user interface development framework,and PyTorch as the deep learning algorithm framework.The system has a friendly user interface,flexible parameter settings,and intuitive inspection results display.