NumLin-Mamba-YOLO: A YOLO object detection algorithm based on Mamba architecture and multi-scale feature optimization

Object detection in complex scenes, especially for small object detection, faces significant challenges in real-time performance and multi-scale accuracy. This study proposes an improved algorithm, NumLin-Mamba-YOLO, designed to enhance detection performance by integrating numerical linear algebra techniques with the Mamba module. The aim of this research is to optimize detection performance through a series of multi-module innovations. The algorithm utilizes Singular Value Decomposition (SVD) and Principal Component Analysis (PCA) in the backbone network for feature dimensionality reduction, denoising, and redundancy elimination, thereby enhancing feature discriminability. It also leverages the linear time complexity advantage of Mamba’s State Space Model (SSM) to model global dependencies. An enhanced feature fusion network improves cross-scale feature correlation, while a decoupled attention detection head is designed to independently optimize classification and regression tasks, improving sensitivity to small objects and local details. Experiments on the Visdrone and PASCAL VOC datasets demonstrate that the algorithm achieves substantial improvements in both detection accuracy and inference efficiency in complex and general scenes, with particular strength in small object detection. The model effectively controls parameters and computational load, providing an efficient solution for real-world object detection applications in intelligent monitoring and autonomous driving.