Abstract: Due to the huge amount of parameters of the YOLOv4 neural network with CSPDarknet53 as the backbone, the detection accuracy and speed will be reduced when it is transplanted to small devices such as mobile phones. In order to improve the detection speed and control the detection accuracy within a reasonable range, this paper proposes to change the original 53-layer neural network to a 15-layer one, and optimizes its clustering algorithm. The K-means++ clustering algorithm is introduced to analyze the data set to generate an anchor box that satisfies the detection conditions. LeakyReLU activation function with a certain slope in the negative interval is used to replace the Sigmoid activation function with vanishing gradients, thereby enhancing the learning ability of the shallow network. At the same time, considering that the center distance and the aspect ratio between the Bounding Box and the Anchor Box have a certain correlation, The corresponding penalty term is added to the original loss function to generate the LCIoU loss function, so that the loss function has a better directionality of the gradient drop during back propagation. Experimental results show that the improved CSPDarknet15 neural network in the VOC2007 data set has an average detection accuracy of 83.94%, and the detection time of a picture is 3 625 ms. Compared with the CSPDarknet53 neural network, the detection speed is increased by 54.43%, which can meet the speed and accuracy requirements of real-time detection of small devices.

Key words: YOLOv4 neural network, K-means++ clustering algorithm, LeakyReLU activation function, LCIoU loss function