Research And Application Of Adaptive Method For Object Detection

With the rapid development of deep learning methods,object detection methods have also developed rapidly.However,in reality,because the experimental data(source domain)and actual application scene data(target domain)have large differences in object appearance,weather,and image quality,the model obtained by training in the source domain is directly applied to the target domain.The detection accuracy is greatly reduced,and the problem of domain drift occurs.The most direct way to solve the problem of domain drift is to re-label and train the target domain image data.However,re-labeling is extremely labor-intensive and time-consuming.The method of solving the domain drift problem without the target domain labeling information is called the domain adaptive method.However,this method is often used in the domain adaptive work of classification tasks.Because the domain adaptive work of object detection tasks not only needs to predict the object category,but also needs to predict the position,it is more difficult,resulting in slow research progress.In order to solve the above problems,this paper studies the problem of object detection domain adaptation in two scenarios,and proposes two methods of object detection adaptation.The main research content includes the following three parts:(1)For the problem of domain adaptation of object detection in the same scene,an adaptive method of object detection based on the fusion adaptive layer is proposed.By adding an adaptive layer after the feature extraction module of Faster R-CNN network,the layer is used to promote The network learns the insensitive feature representations between the two domains and improves the model's adaptive ability.Based on two sets of comparative experiments,it is verified that this method has a performance improvement of about 1.22% to 1.47% compared with other adaptive methods.(2)For the domain adaptive problem of object detection in different scenarios,in order to achieve a strong alignment effect on common features,an adaptive method for object detection based on CycleGAN and presudo label is proposed.CycleGAN is used to transform the source domain data set into an intermediate domain data set,which is then input to the network for training.A preliminary domain adaptive model is obtained,and the model is used to obtain the target domain data set with false label information.Finally,it is fed into the network in turn with the intermediate domain data set for iterative training.The confidence function in the pseudo-label information is used to design the related loss function,constrain the pseudo-label,and maximize the common features between the two domains.Based on a set of comparative experiments,it is verified that the proposed method has a performance improvement of 3.44% ~ 4.21% compared with other adaptive methods.(3)A object detection domain adaptive system is designed and implemented.The system can detect the image data uploaded by the user.After the detection is completed,the detailed detection result can be seen in the browser interface.The system interface is simple,intuitive and easy to operate.