GFSRNet: Lightweight super-resolution for stored wheat quality inspection

High-throughput wheat-kernel quality inspection in warehouse settings depends on reliable image quality; however, cost-constrained imaging hardware and acquisition disturbances can degrade image quality, obscuring kernel morphology and surface details that are critical for quality discrimination and thereby reducing recognition accuracy. To address this issue, we propose GFSRNet, a lightweight super-resolution network that integrates learnable feature distillation with content-adaptive multi-scale spatial modeling. A dynamic multi-scale large-kernel attention module captures contextual information over multiple receptive-field ranges, while channel-wise mean and standard-deviation statistics are used to recalibrate feature responses. Experiments were conducted on the GrainSpace M600 dataset under seven controlled degradation profiles and three upscaling factors (×2, ×3, and ×4). Under the Realistic degradation profile, GFSRNet achieved the highest PSNR and SSIM at all three scales, reaching 26.93 dB and 0.8397, respectively, at ×4. The ×4 model contains 0.329 M parameters and requires 2.898 GFLOPs. Compared with SSIU, WHANet, and UCAN, it reduces the parameter count by 53.3%–73.7% and FLOPs by 54.3%–72.2%. With a fixed ConvNeXt V2 classifier, GFSRNet increased classification accuracy from 36.57% for LR inputs to 78.14% and Macro-F1 from 33.40% to 78.58%. These results indicate that GFSRNet can recover both image structure and task-relevant discriminative information at low computational cost, making it a promising front-end enhancement method for resource-constrained wheat-quality inspection in warehouse environments.