Object Detection With Deep Learning Based On Attention Mechanism And Domain Adaptation

Object detection is an important research content in the field of computer vision.Its main function is to determine the locations of objects in a given image or video frame and classify the objects into specific categories.It is widely used in transportation,national defense and security,industrial production and other fields.With the rise of deep learning technology,object detection methods based on convolution neural network has made great progress and has become the mainstream technology in the field of object detection.Convolution neural network has powerful feature abstraction ability and has made great breakthroughs in various image recognition tasks.However,the convolutional features extracted by convolution neural network are strongly translation-invariant,While ensuring the robustness of the classification model,it can't satisfy the translation sensitivity requirements of the location model in the object detector,resulting in the performance of the object detector can't be further improved.In addition,a large number of labeled training data are needed in the model training process,and in the practical application,the test data and the training data are required to be drawn from an identical distribution.This gives rise to applying the trained model in one scene directly to another different scene,which not only causes the waste of data resources,but also reduces the universality of the model.To counter the problems above,this paper mainly focuses on enhancing the location and recognition ability of the object detection model and improving the domain adaptability of the object detection model.the specific research contents include:1)In terms of the translation invariance of object detection model,a two-stage object detector based on attention mechanism is proposed which is under the inspiration of RFCN model and Class Activation Maps.The attention module is mainly used in the ROI feature extraction process of the two-stage object detection model.In the form of sub-region active attention maps and aspect ratio active attention maps,the translationvariant features of objects are encoded and used to weight and refine the original ROI pooling features,so as to enhance the translation sensitivity of the model and improve the detection accuracy of the object detector.Meanwhile,the use of attention maps makes it possible to reduce the dimension of the ROI features,which can effectively improve the detection speed of the model.2)In order to improve the domain adaptability of the object detector in the process of applying the model trained in the source domain to the target domain detection task,on the basis of the DA Faster RCNN model,we propose to use partial alignment operation and consistency constraint operation to enhance the knowledge transfer between domains and reduce domain shifts.Concretely,partial alignment operation can assist the detection model in extracting robust domain-invariant features by learning from the similar adversarial features and putting less emphasis on features that are not similar between domains.The consistency constraint operation stabilizes the training process and optimize the alignment process by promoting the synchronous alignment of image level domain classifiers and target level domain classifiers.By conducting ablation experiments on several datasets,the experimental results show that the proposed algorithms perform well and effectively improve the detection accuracy and domain adaptability of the object detection model.