Research On Object Detection Based On Attention Mechanism And Neural Architecture Search

The goal of object detection is to determine the spatial position and range of each object instance in the image.As the cornerstone of image understanding and computer vision,object detection is the basis for solving complex or high-level vision tasks,and is applied to many practical business scenarios.At present,the existing One-Stage detectors have many problems,such as insufficient multi-scale feature fusion,features misalignment,and no independent prediction output for different scale features.Therefore,the main research purpose of this paper is to analyze the causes of these problems in One-Stage detector,and propose a targeted improvement plan for each problem of the One-stage detector,and finally improve detection accuracy of the model.First,in view of the insufficient multi-scale feature fusion of the current One-Stage detector,this paper proposes a new feature fusion structure,NFS.Based on the feature fusion structures of FPN and PANet,NFS adds short-cut connections and 3D convolution fusion structures,which makes the semantic information and spatial information of depth features at various scale more abundant,and further improves the multi-scale feature fusion effect,and can bring 1.7 m AP on MS COCO dataset.Then,to deal with the problem that the features misalignment in the One-Stage detector,this paper proposes an improved non-local self-attention module BWNL.Based on the current aggregation of global feature information by the non-local self-attention network,BWNL further introduces the second feature collection.BWNL is applied to the feature alignment process of the one-stage detector,so that the feature map obtains more accurate feature expression corresponding to the visual object.Experiments show that BWNL can significantly improve the feature alignment problem of the One-Stage detector,and benefit the mobel by 0.8 m AP on MS-COCO dataset.Finally,in view of the problem that the head in the One-Stage detector does not use an independent prediction output head module for processing different scale features,this paper proposes to use the NAS to search for independent heads for multiscale feature maps.Experiments show that the new head design obtained by searching can significantly improve the feature extraction ability,can cope with the change of the object feature distribution under the change of the scale of the feature map,and have 1.4 m AP gain on MS COCO dataset.