Response-decomposed residual filtering for joint low-light enhancement and deblurring

Abstract Joint low-light enhancement and deblurring require local corrections that restore visibility and detail together. We propose response-decomposed residual filtering at LEDNet's native dynamic-kernel interfaces. The original generator is retained, while an added route combines a signed kernel-sum increment with four learned zero-sum shape bases. Encoding features predict the sum increment; the same features, augmented by the degraded red-green-blue (RGB) image, predict the shape coefficients. The route operates at three scales and starts from zero output before joint training. Paired fixed-budget comparisons examine the complete model, conditioning source, kernel-sum branch and carrier choice. A prespecified model trained for 60,000 updates is evaluated on 1,800 official-test pairs from 30 videos. Relative to paired LEDNet, the structural similarity index (SSIM) increases by 0.004656 and learned perceptual image patch similarity (LPIPS) decreases by 0.011445, with improvements in 26/30 and 30/30 videos. The peak signal-to-noise ratio (PSNR) difference is -0.045441 dB; PSNR improvement is not established. The route adds 226,880 parameters and approximately 17.7% measured forward latency. These results establish a within-LEDNet SSIM/LPIPS benefit with additional cost; public DarkIR-m remains a stronger system reference under a different training protocol.