Deep Learning Object Detection Based On Attention Mechanism

Object detection studies how to exactly locate and classify objects,which is an important research field in computer vision.As a basic recognition research,object detection has a crucial impact on subsequent studies of pedestrian detection,face recognition,automatic driving and so on.Although the research on object detection has a long history,there are still many key problems that need to be solved urgently,such as the detection accuracy of small objects is generally low,the detection of overlapping objects is difficult,the accuracy of the one-stage object detection algorithm is poor,the balance between accuracy and speed of real-time target detection etc.This paper proposes a deep learning algorithm based on attention mechanism to carry out object detection,the attention mechanism is targeted weighting the extracted features,so that the object detection is more purposeful,the output characteristics are more significant.By designing and training a novel three-branch object detection model(TBANet),including spatial network,semantic network and context network,which is to achieve high quality acquisition of spatial detail information,multilevel semantic information and contextual information,so as to achieve accurate object detection.This paper studies the attention mechanism in depth.Firstly,the research status of attention mechanism,multi-branch network and network backbone is introduced.Then,the theory of convolutional neural network,object detection algorithm and attention mechanism is described.After that,the design of various attention modulesand space pyramid pooling modules is explained,and a comparative experiment is carried out to illustrate the function and effectiveness of the designed modules.The proposed algorithm achieved an accuracy of 83.71% m AP on the PASCAL VOC2007 test dataset,which is better than the most advanced algorithm at present.Finally,the proposed TBANet has been slightly modified to form a semantic segmentation network model,which reached an accuracy of 83.1% Mean Io U on the Cam Vid test dataset,surpassing the current advanced semantic segmentation algorithm and proving the versatility of the proposed algorithm.