Logo detection has a wide range of applications in areas such as brand recognition and intellectual property protection. In order to solve problems of poor detection performance on small-scale logo and inaccurate logo positioning, a logo detection method is proposed based on the YOLOv4 network. Five continuous convolutional layers in the PANet module of YOLOv4 network are replaced by the designed adaptive residual blocks to enhance the utilization of shallow and deep features and fuse features with emphasis and optimize the model training. And the coordinate attention mechanism is used after the adaptive residual blocks to encode channel relationship and long-term dependencies through precise location information, filter and enhance the more useful features from the fused features. The K-means++ clustering algorithm is used to obtain anchor boxes which are more suitable for the logo datasets and assign those to different feature scales. The experimental results show that the mean average precision of the proposed method on FlickrLogos-32 and FlickrSportLogos-10 datasets reaches 88.09% and 84.72%, which is 0.91% and 1.40% higher than the original algorithm, respectively. The performance of the proposed method in positioning accuracy and small-scale logo detection is significantly improved.