基于Swin-Transformer的目标检测算法的研究

Research on Target Detection Algorithm Based on Swin-Transformer

YOLOv4算法作为目前最杰出的目标检测算法之一,已在许多应用中取得了很好的成就,但受限于CNN网络的局限性,其只能在局部区域建立一定联系,而无法与更远的位置建立远程依赖关系。Swin-Transformer因为其独特的自注意力机制,并在任意位置之间均能建立联系。针对以上问题,本文将Swin-Transformer与YOLOv4算法相结合,取长补短,以获得更好地捕捉远程依赖关系的能力。在YOLOv4的基础上,采用Swin-Transformer网络模型作为骨干特征提取网络,并引入了ASFF模块来增强特征提取能力。实验结果表明,在Pascal VOC数据集检测出的精度比标准YOLOv4高出8.1%,证实了该方法的有效性。

The YOLOv4 algorithm,as one of the most outstanding object detection algorithms at present,has achieved great success in many applications.However,due to the limitations of CNN networks,they can only establish certain connections in local areas and cannot establish remote dependencies with further locations.Swin-Transformer,due to its unique self attention mechanism,can establish connections between any position.In response to the above issues,this article combines Swin-Transformer with YOLOv4 algorithm to learn from each other′s strengths and weaknesses,in order to achieve better ability to capture remote dependency relationships.On the basis of YOLOv4,the Swin-Transformer network model is adopted as the backbone feature extraction network,and the ASFF module is introduced to enhance the feature extraction ability.The experimental results show that the accuracy detected in the Pascal VOC dataset is 8.1%higher than the standard YOLOv4,confirming the effectiveness of this method.