Abstract
Instrument reading detection in industrial scenarios poses significant challenges due to reading contour distortion caused by perspective transformation in the instrument images. However, existing methods fail to accurately read the display automatically due to incorrect labeling of the target box vertices, which arises from the vertex entanglement problem. To address these challenges, a novel Quadrilateral Contour Disentangled Detection Network (QCDNet) is proposed in this paper, which utilizes the quadrilateral disentanglement idea. First, a Multi-scale Feature Pyramid Network (MsFPN) is proposed for effective feature extraction to improve model accuracy. Second, we propose a Polar Coordinate Decoupling Representation (PCDR), which models each side of the instrument contour using polar coordinates. Additionally, a loss function for the polar coordinate parameters is designed to aid the PCDR in more effectively decoupling the instrument reading contour. Finally, the experimental results on the instrument dataset demonstrate that QCDNet outperforms existing quadrilateral detection algorithms, with improvements of 4.07%, 1.8%, and 2.89% in Precision, Recall, and F-measure, respectively. These results confirm the effectiveness of QCDNet for instrument reading detection tasks.