Package Number Detection Method Based on Improved YOLOv5 in Complex Environments

Hot metal can not fall to the ground is the most important requirement in the steel plant hot metal scheduling project. The identification of the hanging tank/sitting tank and the packet number of the molten iron tank is the most critical in the process, and the final decision is left to the manual. Due to the chaotic environmental background, complex lighting conditions, high temperature and high pressure, loud noise, heavy dust, non-standard spraying of iron can bag number, serious wear, paint off and other problems, the detection environment is extremely harsh, making conventional methods ineffective, detection results become extremely unreliable, affecting manual interpretation and decision-making, and the production accident rate increases sharply. As a convenient, fast and easy to deploy network detection model, Yolov5 has been favored by many researchers. However, due to the complex environment of our project, the conventional Yolov5 target detection accuracy is poor, the robustness is poor, and false detection often occurs. Although many researchers have proposed different Yolov5 target detection models, however, when the improved model is applied to our complex environment, the results are not improved, and problems still exist. In response to the above problems, This paper proposes an improved YOLOv5 algorithm to address the challenges of detecting the status of molten iron ladles and recognizing packet numbers in complex scenes. The model has been improved in three aspects: ① Replacing the original YOLOv5 convolution module with the GSConv module (Grouped Shuffle Convolution). This significantly reduces the computational load, meeting the real-time requirements of visual detection in molten iron dispatching, and addresses the difficulty of the original model in effectively extracting target features due to poor imaging quality under harsh production conditions. ② Replacing the FPN (Feature Pyramid Network) structure with BiFPN (Bidirectional Feature Pyramid Network). This allows the model to fuse features at different scales, improving the detection and recognition rate of targets under poor imaging conditions. ③ Adding the CoordAtt (Coordinate Attention) module. This enhances the model's ability to focus on different positions on the feature map, further improving the detection and recognition rate of targets in complex imaging conditions. Comparative experiments founded on real data disclosed that, as opposed to other methods, the proffered approach attains superior target detection accuracy with conspicuously reduced parameters and a lightweight model. Furthermore, it manifests outstanding generalization ability, substantiating that our targeted network optimizations are efficacious and can satisfy the operational requisites of real-world applications, presenting considerable practical value.