|
Target detection technology is crucial in the domain of machine vision. Target detection technology, when paired with intelligent robots, enables industrial production to accurately determine the position of the target object. This integration enhances the level of automation in production, leading to improved production efficiency. In order to enhance the accuracy of target recognition, target detection technology relies on a diverse range of object samples. However, target detection models used in industrial production often face challenges such as limited real-time applicability due to their excessive model size and numerous parameters. In order to solve this problem, this paper proposes a method to train a detection model by combining the production of detection samples with digital twin technology, and to lighten and improve the model trained with this class of synthetic samples, we propose an improved target detection algorithm based on synthetic samples of yolov8n, called fast_yolov8. In the fast_yolov8 model, Dualconv structure and Adown structure are introduced in the backbone part, and AKConv is introduced in the Head part as a lightweight detection head, and the introduction of these three lightweight structures significantly reduces the computational cost and the number of parameters while guaranteeing the detection accuracy of the model. The experimental results show that compared with the pre-trained yolov8n model, the model parameter quantity of fast_yolov8n based on the synthetic sample set is reduced by 25%, the floating-point operations (FLOPs) are reduced by 23%, and the detection accuracy of mAP on the synthetic sample set is improved by 1.3%, which realizes the lightweighting of the model and maintains the detection accuracy of the model.
|
