FDFNet: An efficient detection network for small-size surface defect based on feature differentiated fusion

In industrial production, real-time detection of steel surface defects is a crucial factor in quality assurance. Furthermore, steel surface defects are diverse and complex, particularly when additive manufacturing of metallic structures have been widely used in the industries. They are easily disturbed by background interference. The current defects detection algorithm requires further enhancement in terms of speed and accuracy. This work investigates the problem of fast and high-precision steel surface defects detection by a lightweight inspection model, FDFNet, based on YOLOv9. First, for the contrast between the defects and the background, a Contrast Limited Adaptive Histogram Equalization (CLAHE) data enhancement strategy is proposed. Second, a SPace-to-Depth Convolution (SPD-Conv) is constructed in the backbone, which retains more texture information and reduce the model parameters. Additionally, a Differentiated Fusion (DF) module is designed at the neck to highlight both the consistency and heterogeneity of feature maps across disparate scales. Finally, the findings of the experiment conducted on the data set of NEU-DET show that the proposed defects detection algorithm can improve the detecting speed and accuracy compared to those of the existing approaches including YOLOv9. To sum up, the proposed model demonstrates an optimal balance between detection efficiency and accuracy.