EdgeVisionPlus: A Visual Recognition Model to support Edge Deployment in Manufacturing Application Scenarios

The manufacturing industry can enhance the efficiency of research and development, production, logistics, and operation through visual recognition of scenes. The implementation of visual recognition in manufacturing scenes faces challenges such as significant data dependency, high computational power requirements, and high operation and maintenance costs. Therefore, studying how algorithms can quickly adapt to changes in segmented scenes, reduce training computational power requirements, and support flexible deployment on cloud-edge-end, has become an important research topic in the industry. This paper proposes a classification algorithm framework for behavior recognition based on the basic vision model and downstream task fine-tuning two-stage method. Experimental results demonstrate that, compared to traditional single-task fine-tuning methods, this framework achieved a 10.8% and 6.8% improvement in average precision (AP) on two small-scale behavior recognition datasets, respectively, while using 1/4 computational power hardware and 1/8 training time. Overall, the proposed method exhibits characteristics such as low data sample dependency, low computational power requirements, high training speed, and high scalability based on a shared basic model.