Research On Object Detection Algorithm Based On Feature Fusion And Attention Mechanism

Object detection is one of the important research directions in the field of computer vision,which is widely used in industrial production.In the detection task,for an input picture,the detection algorithm distinguishes the foreground and background of the picture by learning,then separate the interested object from the background,judge the position and category of the object in the end.However,due to the small area and low resolution of small objects,the detection accuracy of small objects can't achieve the desired results.To solve the problem of small object detection,this paper makes an improvement based on SSD model,and realizes a higher precision target detection algorithm through multi-scale cross-layer fusion,attention mechanism,channel Hierarchical-Split fusion and more effective nonmaximum suppression algorithm.The main research contents of this paper as follows:(1)The basic network model extracts feature information through the backbone network,and six features with different scales are obtained.On this basis,this paper fuses the information of two feature maps with different scales through cross-layer connection,and obtains richer features.Low-level features have high resolution and contain more position and detail information,but they have lower semantics and more noise due to less convolution,while high-level features have stronger semantic information,but their resolution is very low and ability of perceive details is poor.This paper studies the existing multi-scale feature fusion methods,and proposes a cross-layer fusion method to realize more effective information interaction between feature layers.(2)According to the characteristics of attention mechanism,In this paper,the channels are weighted after cross-layer fusion,so that the network model pays more attention to the key channels containing objects,while ignoring the image areas that have less influence on the detection effect.Because each pixel in the image has different importance to the detection result,this paper introduces attention mechanism,which weights different channels according to their importance,and effectively improves the feature extraction of the effective area of the image.(3)This paper studies the influence of receptive field on object detection performance.After many convolution operations,the pixels on each layer of feature map corresponding to the area size of the input image is different,in other words,the size of the receptive field is different.The larger the receptive field,the more global information contained on the pixel of the output layer.Unlike previous object detection algorithms which used dilated convolution to increase receptive field,this paper uses the latest channel Hierarchical-Split fusion method to obtain features with larger receptive field and more information,without increasing computational complexity.In addition,Soft-NMS is used instead of NMS in the detection part to alleviate the problem of missing detection and improve the recall rate of detection algorithm.Experiments show that the proposed object detection algorithm makes the best of the effective image information,obtains better features,and achieves higher detection accuracy and recall rate.