Conveyor belt foreign object detection method based on improved YOLOv11 and ESRGAN

Traditional conveyor belt object detection methods often lack robustness and adaptability under challenging conditions such as low-light and low-resolution environments. This study proposes an improved detection method specifically designed for conveyor belt environments, built upon the YOLOv11 object detection framework. A custom dataset was created to support foreign object detection on factory conveyor belts. To overcome the low resolution of image recognition, the Enhanced Super- Resolution Generative Adversarial Network (ESRGAN) was employed to improve the input image clarity. Additionally, to enhance the performance under low-illumination conditions, several architectural improvements were embedded in the YOLOv11 framework, leading to the proposed Conveyor Belt Foreign Object Detection (YOLOv11-CBFD) algorithm. These enhancements included an optimized upsampling module, integrated attention mechanisms, a modified convolution module, an improved loss function, and a modified convolution module. Experimental results demonstrated that the proposed YOLOv11-CBFD algorithm significantly enhanced the accuracy of foreign object recognition. Based on a dataset collected from a factory conveyor belt, YOLOv11-CBFD achieved an accuracy of 86.1%, a recall of 86.7%, an [Formula: see text] of 89.1%, and a model size of only 2.17 M parameters. Compared to the original YOLOv11n model, the proposed method reduced the parameter count by 16.2% while demonstrating no significant degradation in recognition capabilities. In terms of computational efficiency, the optimized architecture demonstrated a 12.4% increase in the number of frames per second when deployed on a Jetson Orin NX-embedded AI computer. Field experiments conducted in industrial inspection scenarios validated the practical effectiveness of the system, demonstrating continuous operation over 48 h under real-time constraints (average latency <33 ms/frame), while consistently maintaining an accuracy of 86.1% across multiple deployment cycles. The experimental results highlight the ability of the model to effectively balance computational efficiency and detection performance on embedded AI platforms.