Addressing the needs and challenges of multi-site feature detection for knee arthritis, this paper proposes an improved lightweight network model, MMG_YOLOv5s. This model aims to solve the problems of complex structures and massive parameter counts in traditional detection models, which hinder their deployment on edge devices. The M(1) and M(2) modules in the MMG_YOLOv5s model enhance the YOLOv5s network by adjusting its backbone and neck structures, decreasing the count of network parameters and compressing the model's size. This, in turn, improves computational efficiency and creates conditions for deployment on resource-constrained devices. The data for this experiment is sourced from the public dataset OAI and annotated under the guidance of professional doctors to construct our own dataset. The experimental outcomes indicate that the recognition accuracy of this model achieves 80.8%, representing an 11% enhancement in comparison to the original YOLOv5s model. More importantly, its parameter count is reduced by 71.27%. This improvement provides a solid foundation for the subsequent embedded deployment of the model into portable or mobile devices, promising to advance the intelligent and convenient process of knee arthritis detection.