DPALSR: Dual Prior-Aware Lightweight Super- Resolution Network

Lightweight single-image super-resolution (SISR) requires effective local structural representation and broader contextual modeling under a restricted computational budget. We propose DPALSR, a Dual Prior-Aware Lightweight Super-Resolution Network that conditions successive attention-query representations with complementary directional cues. Learnable Constrained Laplacian Attention (LCLA) derives local directions from the gradient of a learnable zero-sum structural response, whereas Fourier Neural Operator Attention (FNOA) derives coarser within-window contextual directions through compact spectral processing. These cues provide structured inductive biases for learned feature interaction. On five standard SISR benchmarks, DPALSR achieves a competitive reconstruction accuracy–complexity trade-off. At ×4, it obtains 27.01 dB/0.8128 on Urban100 and 31.52 dB/0.9203 on Manga109 using 887K parameters and 59.8G reported FLOPs. Relative to the published MambaIRv2-Light operating point, these PSNR values are higher by 0.19 and 0.28 dB, respectively, with approximately 20.9% fewer FLOPs.