An Efficient Super-Resolution Network Based on Spatial-Frequency Loss for Precision Measurement

Two-dimensional visual measurement is a noncontact technique that is widely applied in industrial fields due to its high precision and efficiency. The resolution of measurement images is a critical factor that directly determines the upper limit of measurement accuracy. Currently, improving image resolution through super-resolution (SR) reconstruction using neural networks has proven to be an effective approach for enhancing measurement accuracy without incurring high hardware costs. However, existing SR methods typically rely on extracting local features with a limited effective receptive field (ERF), which hinders their ability to balance computational cost and performance. Moreover, these methods often lack the capability to capture high-frequency (HF) information such as corners and edges, which are crucial for precision measurement and edge detection. To address these limitations, this work proposes a novel SR model for precision measurement called B-Biformer-SR. First, the model enlarges the ERF and reduces complexity through downsampling, which is not common in SR tasks. The model also incorporates the proposed mixed feature aggregation block (MFAB), which enhances SR performance by enabling multiscale interaction between spatial and channel features. Moreover, edge loss and wavelet loss are designed to capture the HF details and better preserve the geometric features essential for precision measurement and edge detection. Extensive experiments on public and custom datasets demonstrate that B-Biformer-SR, not only achieves promising results on peak signal-to-noise ratio (PSNR) and produces better visual perception, but also exhibits the lowest errors in precision measurement and edge detection tasks. In particular, the model reduces the average measurement error for small objects by 21% compared to the suboptimal baseline model. Code is available at https://github.com/RayTan183/B-Biformer-SR