Research On Object Detection Technology Of Image Based On Convolutional Neural Network And Motion Characteristic

In recent years,Artificial Intelligence(AI)has gradually become a popular technology field,and many industries have successively proposed the concept of "AI+XX",which has greatly promoted the development of AI.Object detection of image is an important development direction in AI.It is of great significance to study object detection technology of image.It can improve the combat effectiveness and intelligence level of troops in the military and its related applications cover clothing,food,housing and transportation in civilian use.Based on summarizing the research status and analyzing the principle of object detection of image,this thesis has carried out research from two aspects: object detection of static image and object detection of sequence image.The main work of this thesis includes:1.From object detection of static image and object detection of sequence image,the research status in object detection of image at home and abroad was combed and summarized,and the current research problems were extracted.This thesis systematically analyzed principles of CNN-based obejct detection algorithms of static image represented by Faster RCNN,YOLOv3 and motion-based object detection algorithms of sequence image represented by frame difference method.On the basis of detailed introduction and analysis of CNN's forward layer-by-layer operation and output characteristics,post-processing of the inference process,algorithm performance evaluation method,the characteristics of two-stage method and one-stage method,frame difference method and moving object area enhancement method,this thesis proposed a research framework.2.Aiming at the problem of how to quickly and accurately realize object detection and recognition of more multiple multi-value attributes with small differences in static image,an end-to-end rapid object detection and attribute recognition algorithm was proposed on the basis of CNN-based static image end-to-end one-stage method YOLOv3.Considering the existence of dependency relationships among various attributes,the attributes were divided into three types: main attribute,dependent attribute and general attribute through planning attributes.A method for determining the value of dependent attribute guided by main attribute was proposed.That is,in the training process,a higher penalty coefficient was assigned to the values that were unlikely to occur in dependent attribute and the classification loss of main attribute in classification loss function.During the inference process,the range of values of dependent attribute was determined by the predicted value of main attribute.Since multiple values of multiple attributes wound be combined to obtain many categories,it was not conducive to NMS processing.Therefore,it was proposed to determine the predicted position of the object based on the location confidence of the predicted object box containing real object as the NMS ranking basis,and the predicted values of various attributes of the object based on the voting method.In addition,top-1 positioning & recognition accuracy and top-3 positioning & recognition accuracy were proposed to measure the algorithm accuracy.3.In order to solve the problem of how to detect small moving objects in sequence image with a low missing alarm and false alarm,based on the motion characteristic of objects,especially the track formed among frames,a small moving objects detection algorithm based on track-related was proposed.The detection of small moving objects was decomposed into preliminary detection of suspected objects and filtering of false objects in suspected objects.Due to the problem of threshold simplification in the frame difference method,a binary threshold setting method for fusion of regional texture features and difference probability was proposed.This method adaptively determined the threshold based on inverse difference moment in each divided sub-region,which could lower missing alarm to a certain extent.Based on the threshold setting method above,frame difference method and moving object area enhancement method,the preliminary detection of the suspected objects using the center position marker was realized.To further lower the false alarm,the false objects in suspect objects were filtered by calculating the relevant parameters of the suspect objects' track associations and setting double thresholds.4.To solve the problem of how to further determine the specific categories of small moving objects and correct their positions in sequence image,on the basis of proposed CNN-based two-stage algorithm of small moving objects recognition,a small moving object detection and recognition system under dual field of view was constructed.The system first registered the imaging of the smaller focal length lens and the larger focal length lens to establish the correspondence between the pixel positions of the common area imaging in the dual field of view formed by the dual lenses.Then,based on the correspondence established above,the region proposals that might contain small moving object were generated in larger focal length lens imaging.Finally,the proposed recognition algorithm was used to recognize the small moving object and further correct their position in each region proposal.In addition,the main operation time in recognition process was analyzed,and the time complexity of each main operation was analyzed to illustrate the effectiveness of the speed-up method and strategy.The interface design,hardware composition,actual operating effect,module composition and coding implementation were introduced in detail.