Label Defect Detection Based on Improved CenterNet

Defects in labeling represent a prevalent issue that significantly impacts product quality during the manufacturing process. In response to the challenges associated with the detection of label stain defects, this research introduces an enhanced version of the CenterNet algorithm, termed Improved CenterNet. This novel approach predicts the centroid of the object, the diagonal length, and the sine value, which collectively contribute to the optimization of the loss function. This innovation effectively mitigates the problem of disproportionately large loss values, facilitating the accurate estimation of the object's dimensions. When compared to the conventional CenterNet, the Improved CenterNet algorithm not only preserves real-time performance but also markedly enhances the precision and robustness of the detection process. By integrating a sine value prediction mechanism, the algorithm demonstrates a heightened capacity to discern stains against intricate background contexts. Empirical findings indicate that the Improved CenterNet algorithm yields superior performance on a dataset of label stain defects, fulfilling the industrial demand for high-efficiency and high-accuracy defect detection in the manufacturing sector.