LEDNet gains a residual branch: structural similarity up 0.0047, PSNR does not follow
Synopsis
At LEDNet's existing dynamic-kernel interfaces, LEDNet-P3 adds a residual branch that splits the kernel increment into a signed kernel-sum coordinate and four zero-sum shape bases, with the degraded RGB image feeding only the shape coefficients. On 1,800 official LOL-Blur test pairs, the paired LEDNet comparison shows structural similarity up 0.004656 and learned perceptual patch similarity down 0.011445, while the PSNR difference is −0.045441 dB, so a PSNR improvement is not established. The branch adds 226,880 parameters and about 17.7% measured forward latency.
Interpretation
Low-light blurred images need visibility and spatial detail recovered together, and LEDNet-P3 adds a learnable increment to the convolution kernel at three native filtering scales while keeping the original generator. Previously LEDNet predicted a 5×5 kernel per location directly from encoding features, leaving the increment's response to neighborhood means and local variation implicit. On 1,800 official LOL-Blur test pairs from 30 videos, the prespecified model raises structural similarity by 0.004656 and lowers learned perceptual patch similarity by 0.011445 relative to paired LEDNet, improving in 26/30 and 30/30 videos respectively.
The increment is split into two parts: a signed kernel-sum coordinate carries the response to a constant neighborhood, and four zero-sum shape bases carry the response to local variation. A 1×1 convolution on encoding features predicts the sum increment, while the same features added element-wise to a bilinearly resized degraded RGB image predict four shape coefficients per channel through another 1×1 convolution. Each of the four 5×5 bases has its own mean over the 25 taps subtracted before position-wise mixing, so the shape term does not respond to a constant neighborhood; the three scales use 128, 128 and 64 channels, the additions total 226,880 parameters, and training starts from zero output.
Perspective
The findings come from 1,800 official LOL-Blur test pairs, with models fixed before test access under a paired protocol of 60,000 updates, batch one and 153 videos; the comparison is against LEDNet trained under the same protocol, while public DarkIR-m and LEDNet have different training histories and serve only as separate system references.
The mean PSNR difference is −0.045441 dB, with 16 videos improving and 14 worsening and saved per-video differences ranging from −2.150967 to +2.420956 dB, so a near-zero overall difference does not mean negligible change in every video. Post-hoc NIQE and LOEhigh rise relative to LEDNet while LOElow falls, so a broader restoration advantage is not established; the DarkIR adaptation also shows no established advantage over ordinary residual filtering.