AS-YOLO: An Improved YOLOv4 based on Attention Mechanism and SqueezeNet for Person Detection

With the development of deep learning in the field of object detection, a series of detection algorithms and methods are proposed to improve their performance. YOLOv4 is superior to the early algorithms in detection accuracy and speed. In order to get excellent performance in person detection, the AS-YOLO is proposed based on the YOLOv4 network. An attention mechanism is introduced in YOLOv4 to enhance the detection capability. Residual structure and SqueezeNet structure are also used as reference to deepen the network and reduce parameters. In addition, the activation function of backbone network is replaced by LeakyReLu to accelerate the detection speed. Compared with YOLOv4, the proposed network can decrease the parameters by 20.49% and increase the average precision by 2.02% on INRIA person data set, and the detection speed is also improved by 16.63%. This proves the validity of the proposed AS-YOLO.