Research On YOLOv4-tiny Objection Detection Network Integrated With Transformer

The YOLOv4 model is one of the models in the field of objection detection.At present,it has achieved good detection results in objection detection,but there is a problem of excessive calculation.YOLOv4-tiny is a lightweight model of the YOLOv4 model,which has the advantages of fast computing speed and small structure,but the detection accuracy of the YOLOv4 model has declined.In order to improve the detection accuracy of the YOLOv4-tiny model,this paper conducts in-depth research on the YOLOv4-tiny model,and proposes a fusion loss function BIOU loss and a YOLOv4-tiny model that integrates the OD-Trans(Objection Detection-Transformer)module,and finally obtains The detection accuracy of the YOLOv4-tiny-BODT model is improved.The main work of this paper is as follows:(1)In view of the insufficient feature extraction of the data set in the YOLOv4-tiny model,the missed detection of occlusion when there are many detected objects,and the low precision caused by the CIOU loss function not taking into account the balance of the sample,it is proposed to use the CIOU loss The function is fused with the BCE(Binary Classification Balanced Cross Entropy)loss function,which balances the positive and negative samples,solves the problem of gradient disappearance,and obtains the best effect through parameter adjustment.On the PASCAL VOC2007 data set,various indicators were compared with other network models.The experimental results show that:the YOLOv4-tiny model with the BIOU loss function has increased the average detection accuracy of the YOLOv4-tiny model by 0.32%,effectively improving the Accuracy of object detection.(2)In view of the fact that the detection accuracy of the YOLOv4-tiny model is not high enough compared with the current objection detection field,this paper adds an ODTrans module to the network structure to further improve the reasoning speed.The multihead self-attention mechanism and MLP perception layer in the OD-Trans module enrich the feature extraction of the network model for pictures,and the position encoding module makes the model no longer need fixed-size input.Comparing various indicators with other network models on the PASCAL VOC2007 dataset,the experimental results show that the average detection accuracy of the YOLOv4-tiny model integrated with the OD-Trans module is 2.67% higher than that of the YOLOv4-tiny model.(3)Ablation experiments were carried out on the proposed improved YOLOv4-tiny model.Comparing the YOLOv4-tiny model,the YOLOv4-tiny-BIOU model with the fusion loss function,the YOLOv4-tiny-ODT with the OD-Trans module,and the YOLOv4-tiny-BODT model with the BIOU and OD-Trans fusion,the average detection accuracy is respectively They are 91.37%,91.69%,94.04% and 94.15%.The experimental results show that the average detection accuracy of the YOLOv4-tinyBODT model proposed in this paper is 2.78% higher than that of the YOLOv4-tiny model.Figure [38] Table [5] Reference [70]...