FBML-YOLOv5: A Small Target Detection Model based on Fastenet and Attention Mechanism

As small target detection requires identifying and localizing objects of small size in an image or video. Compared with large target detection, small target detection requires less information to be processed and is more susceptible to factors such as noise, background complexity and resolution. Aiming at these problems, this paper proposes an improved YOLOv5-N model, named FBML-YOLOv5, specifically for small target detection. Firstly, the C3 module is improved and the C3 Fast module is proposed. It effectively reduces the number of parameters, and improves the model feature expression ability and nonlinear expression ability. Then, for the neck part, a feature adjustment layer for small target detection is added for high-resolution feature map processing, and in order to strike a balance between performance and complexity, a lightweight hybrid local channel attention (MLCA) module is incorporated into this module to improve the performance of the object detection network, which is capable of simultaneously integrating the channel information with the spatial information, as well as the local information and the global information to improve the representation of the network. Experimental results on the VisDrone2019 dataset show that the FBML-YOLOv5 proposed in this paper outperforms other models for small object detection. Compared with YOLOv5-N, the parameter decreases by 0.09%, map@50 Increase 4.9%.