Research On Small Target Detection Model Based On Optimized YOLOv5

The wide application of deep learning technology in the field of computer vision has led to the rapid development of target detection technology.The detection accuracy of the current detection model for normal size targets has improved rapidly,but the detection accuracy of small targets is far lower than that of large and medium targets..In real-life scenarios,the detection of small objects is also crucial.For example,the successful detection of small foreign objects such as nails on the airport runway can reduce the incidence of aviation accidents.Small target detection has difficulties such as low target resolution,large interference in complex scenes,and lack of large and complete small target detection data sets.The main problem that hinders the development of small target detection technology at this stage is that it is difficult to take into account the detection speed while improving the detection accuracy;the transferability of the detection technology is poor,and most of the existing small target detection technologies are designed according to specific scenarios,such as small face detection,cigarette detection and so on,the adaptability to scene changes is not strong.In view of the above problems,this paper proposes a small target detection model based on optimized YOLOv5-AF-O-YOLOv5 model(Attention-Feature Fu Sion-Optimization YOLOv5).The model mainly achieves three improvements: changing the feature extraction structure of the Backbone part of the YOLOv5 benchmark network,performing the original network feature extraction operation in advance,extracting the target features from the first C3 module of the Backbone structure,and horizontally integrating it into the Neck layer On the feature layer of the same scale;the CA attention mechanism is introduced after the SPPF structure,which is intended to reconstruct the feature weight of the detection target and the background information;the context feature fusion structure is added to the head end of the YOLOv5 network,which uses3 head ends to be used.The feature maps of the predicted target are fused,and the transposed convolution is used to transform the width and height scales of the deepest and sub-deep features,and the channel scales of the two are set to half of the shallow feature channel scale,which is used as context information to stitch with the shallow features.Perform feature fusion.The AF-O-YOLOv5 model achieves m AP of 41.5% on the aerial data set Vis Drone2019-DET verification subset,which is 8.5% higher than the YOLOv5 benchmark network;m AP reaches 32.6% on the test subset,which is 3.5% higher than Faster R-CNN.