Petrochemical Equipment Detection by Improved Yolov5 with Multiscale Deep Feature Fusion and Attention Mechanism

Petrochemical equipment detection technology plays important role in petrochemical industry security monitoring systems, equipment working status analysis systems, and other applications. In complex scenes, the accuracy and speed of petrochemical equipment detection would be limited because of the missing and false detection of equipment with extreme sizes, due to image quality, equipment scale, light, and other factors. In this paper, a one-stage attention mechanism-enhanced Yolov5 network is proposed to detect typical types of petrochemical equipment in industry scene images. The model considers the advantages of the channel and spatial attention mechanism and incorporates it into the three mainframes. Furthermore, the multiscale deep feature is fused with a bottom-up feature pyramid structure to learn the features of equipment with extreme sizes. Moreover, an adaptive anchor generation algorithm is proposed to handle objects with extreme sizes in a complex background. In addition, the data augmentation strategy is also introduced to handle the relatively small and extremely large sample and to enhance the robustness of the fused model. The proposed model was validated on the self-built petrochemical equipment image data set, and the experimental results show that it achieves a competitive performance in comparison with the related state-of-the-art detectors.