Lightweight detection method for industrial gas leakage based on improved YOLOv7-tiny

In the scenario of industrial gas leakage, traditional object detection models face challenges such as high computational complexity, large parameter count, and slow detection speed, making it difficult to deploy them on terminal hardware devices with limited computational resources. Meanwhile, existing lightweight object detection models also encounter difficulties in balancing detection accuracy and real-time requirements. To address this issue, this paper proposes a lightweight object detection algorithm, P-YOLOv7-TFD, based on the YOLOv7-tiny model. Firstly, the YOLOv7-TFD algorithm is constructed, which utilizes FasterNet as the backbone network, reducing the computational complexity and improving feature representation capability through the application of rapid feature fusion and efficient upsampling modules, while maintaining accuracy and enhancing detection speed. Secondly, the network's head layer is reconstructed based on the diverse branch block module to enrich the diversity of feature space, thereby improving model performance without sacrificing inference speed. Finally, the YOLOv7-TFD model is pruned using the Network Slimming channel pruning algorithm to obtain the P-YOLOv7-TFD model, further reducing the model's parameter count. Experimental results show that, compared to the original YOLOv7-tiny model, the average precision of the P-YOLOv7-TFD object detection model decreases by only 1.6% on a self-built dataset, while the parameter count and computational load decrease by 75.7% and 71.8% respectively. The computation inference time is decreased by 88.3%, and the model weight size is only 4.4 MB, representing a reduction of 62.4%.