Research On Object Detection Based On Attention Model

Object detection,as a basic technology in computer vision,aims to locate and classify all instances of predefined object class in images.At present,although some mainstream object detection algorithms have achieved good results in the accuracy and speed of detection,there are still problems such as insufficient utilization of object feature information and lack of consistency between classification confidence and localization accuracy in detection results.The attention model can assist the network to focus on the object areas and improve the network utilization of object information.Therefore,in order to solve the above problems,the paper studies the object detection based on attention model,and proposes Non-Maximum Suppression(NMS)optimization algorithm based on attention map and object detection optimization algorithm based on spatial attention map.The main work of thesis can be summarized as:NMS algorithm is a key step to achieve object location,but the algorithm will suppress some bounding boxes with high localization accuracy but low confidence.To solve this problem,the paper proposes a non-maximum suppression optimization algorithm based on attention map.By back propagation of the object high-level semantic information to reconstruct features of the object,and an object attention map is generated by weighting it.Then the interest probability of the detect bounding box(which refers to the probability accumulation value of the regions contained on the normalized attention map)and the classification confidence are weighted to obtain the interest score,which can be used as the ranking key of the NMS algorithm to obtain the best bounding box of the object.Experiments based on PASCAL VOC2007,PASCAL VOC2012 and MS COCO data sets verified the effectiveness of the algorithm.(2)The bottom-up image information contains regional features of object and background,which can focus on the relationship between the object area and other areas;the top-down high-level semantic information is the mapping of the bottom image information to the top-level output information,which can focus on the object typical characteristics.Based on this,an optimization algorithm of object detection based on spatial attention map is proposed.The algorithm combines bottom-up image information with top-down high-level semantic information to expand the coverage area of the object in the image so as to extract richer object features.In the bottom-up detection process of Faster R-CNN network,the Transformer attention module is introduced to enhance the connection between the regions in the image,and generates a map of object interested area,then it is combined with top-down object high-level semantic information to generate a complete object attention map.Finally,a spatial attention map is generated based on the object attention map,which is used to optimize the classification and regression of the object feature information.The experimental results on PASCAL VOC2007,PASCAL VOC2012 and MS COCO datasets show that the algorithm is effective for improving the detection accuracy of the object.