A small target detection algorithm based on an improved YOLOv8 detection head

Order to improve the detection accuracy for target detection, an improved algorithm based on YOLOv8s network architecture is proposed. In this paper, the network architecture is mainly optimized for the detection head. Firstly, the shared convolution is introduced into the detection head; secondly, the planning layers of the convolution are replaced by the group normalization; furthermore, the interaction between the classification and localization branches of the detection head is strengthened with reference to the idea of task alignment; lastly, in order to solve the problem of weakening or even disappearing of the small target features due to deepening of the number of layers of the network, a small-target detection layer is added, and the fusion of the front-side network features and the deeper network features is performed, so as to further preserve the small-target features. The small target features are further preserved by fusing the front network features with the deep network features. The experimental results show that on the VisDrone dataset, the average accuracy of the original YOLOv8s detection is 37.2%, and after the improvement, it reaches 49.5%, which realizes the improvement of the detection accuracy