Design And Implementation Of Object Detection Algorithm Based On Cross-layer Attention Mechanism And Multi-scale Perception

Object detection is one of the basic tasks of computer vision.Its purpose is to detect task objects in images or video streams.It has good application scenarios in the fields of security and transportation.For the current mainstream object detection algorithms,most of them have a problem,they are very sensitive to the scale change range of the object of interest to be detected,and the detection object with a larger scale change range is less robust,especially for small objects..In order to better promote the application of object detection technology in the above-mentioned fields,designing a detection algorithm that can deal with small objects and a large range of object scale changes has become a research focus in this field.Based on the SSD algorithm,this paper proposes a object detection algorithm based on cross-layer attention mechanism and multi-scale perception.The algorithm uses spatial attention mechanism for low-level features and channel attention mechanism for high-level features to further refine the feature information of the object area,generate effective object features,and for high-level feature maps with rich semantic information,channel attention can be model the relevant information between channels to improve the algorithm's object classification ability.In addition,the detection algorithm also introduces a feature pyramid and a multi-scale perception module to improve the robustness of the network to object scale changes,and adopts feature adaptive fusion in the multi-scale perception module instead of simply performing equal-scale features.To a certain extent,helmet detection belongs to the category of small-scale detection.Therefore,for helmet detection,we also propose a fast and effective anchor frame adaptive adjustment method,which can adaptively adjust each anchor frame according to the size of the feature map.The scale distribution of the layer anchor frame,which is conducive to the detection of small objects on the low-layer feature map.Experimental results show that the algorithm has an average accuracy of 80.5% on the VOC 2007 test set,and an average accuracy of 88.1% on the open-source helmet data set GDUT,achieving high detection accuracy.