Research And Application Of Video Target Detection Based On Deep Learning

In recent years,the target detection method for static images based on deep learning has made significant progress,but for video target detection,if the static image target detection network is simply applied to video detection,the target may not be detected on a single frame due to the motion blur or the out-of-focus blur.In addition,since the image change between adjacent video frames is slow,if deep image neural network extraction is used to extract features for each frame of image,there is a case where computational redundancy is performed.Therefore,a network framework for video object detection based on deep feature flow is proposed.The core of the algorithm is to extract features by using convolutional networks only in sparse key frames.In non-key frames,their features are propagated by the features of adjacent key frame through the optical flow field.This way can greatly reduce most of the calculations and improve the efficiency of video target detection.On the basis of video target detection network based on deep feature flow,this thesis carries out research on the detection of teeth video.The main work of this thesis is as follows:1.Based on the video frame detection network framework based on depth feature stream,an improved adaptive key frame algorithm is proposed.Key frames are selected according to the two evaluation criteria of definition and motion displacement,and the key frame is used to propagate the feature map before and after instead of only backward.This method can reduce the jitter problem and improve the detection accuracy2.An improved target channel algorithm is proposed,which uses the way of combining IoU and optical flow to connect the same target object corresponding to different frames of the video stream.This method can improve the accuracy of prediction compared with the method of calculating only IoU.3.Created a teeth video target detection data set and a teeth video segmentation data set.In addition to using data enhancement to extend the data set,an optical flow method is also used to extend the annotation of the sparse frame of the dental video object detection data set to all video frames,by which way all video frames are marked with rectangular labels;4.Apply the deep learning video target detection network to the teeth video to locate or segment each tooth in the video for the first time.It is proved by experiments that the deep learning target detection network can be applied to tooth detection and segmentation,and can achieve good results,and the improved adaptive key frame algorithm can balance the detection speed and accuracy well.5.Developed an orthodontic application software based on the deep learning video target detection network.Through the improved video object detection algorithm based on depth feature stream,this software can detect the teeth in the video and use the improved target channel algorithm to determine the position of the teeth to be replaced in each frame to replace the beautified tooth model.And the static image and dynamic video effect after orthodontics can be displayed to the patient on the tablet.