| With the development of chip technology,more and more devices are used to obtain optical images and videos,such as cameras,drones,mobile phones,etc..How to obtain interesting visual targets from a large number of images and videos is of great significance.Among them,image object detection needs to identify and locate specific types of targets in the image,which can be used as the basic technology of video object detection.Therefore,the paper mainly studies how to use deep learning and recurrent sequence learning to detect and recognize objects in videos and images.A method for image object detection is first proposed,and then based on it,two video object detection methods are designed.1.A method based on pyramid dilated convolution network for image object detection is proposed.This method is improved based on single shot multi-box detector(SSD).To solve the problem that the expression ability of shallow feature maps in SSD is not strong,we design a multi-scale pyramid dilated convolution module to learn the multiscale information of objects for the shallow feature map.In addition,a dilated residual block is proposed based on the traditional residual block to further extract features,thereby generating multiscale feature maps for detection.The proposed method is experimented on the optical image dataset VOC and the remote sensing image dataset NWPU VHR-10.Through the designed module,the accuracy is gradually improved.In addition,compared with the existing onestage object detection algorithms SSD,YOLO v2,two-stage object detection algorithms FPN,RFCN,etc.,our method achieves the higher accuracy,indicating the effectiveness of the method.2.A method based on the tubelet proposal generation is proposed for video object detection.This method aims to address missed and false detection problems after using the single frame image object detector to detect each frame of the video throuth applying the designed tracking module and correction module.Among them,the tracking algorithm is used to generate the tracking tubelet proposals,which can capture the object missed due to the motion blur or occlusion.In addition,the designed false positive analysis strategy can delete false detections,because it can associate the contextual relationship between frames,improving the accuracy of video object detection.The method is experimented on Vis Drone and Image Net VID datasets.Through the added modules,the recall rate and accuracy of objects are improved.Moreover,compared with other post-processing-based video object detection methods Seq-NMS,T-CNN,end-to-end FGFA,our method achieves the better precision,which verifies the effectiveness of the proposed method.3.A method based on recurrent sequence learning and pyramid dilated convolution is proposed for video object detection.This method abandons the multi-stage video object detection framework with slow detection speed and builds an end-to-end detection network based on recurrent sequence learning.The long-short term memory module is designed to learn the timing correlation between frames,thereby enhancing feature representation of each frame.Moreover,to improve the detection accuracy of multiscale targets,the multi-scale pyramid dilated convolution module is used to learn the multiscale information of objects on the shallow features.This method is experimented on Image Net VID and Vis Drone video datasets.Compared with other video object detection algorithms DFF,FGFA,STMN,etc.,the higher accuracy is obtained,proving the superiority of the proposed method. |
