Table 1 Main parameter settings of the improved Canny–EdgeConnect–DPSRGAN framework.

Role

Preprocessing (non-trainable module)

Trainable network (Stage 1)

Trainable network (Stage 2)

Network/Structure

Bilateral filtering + Otsu adaptive thresholding

Two-stage generation network (edge generator + image completion network)

Based on SRResNet + RRDB modules + U-Net discriminator

Key Parameters

σs = 2.0; σr = 25.0; filter = 5 × 5; gray levels = 10; image resolution = 256 × 256

Batch size = 8; input size = 256 × 256; learning rate = 1 × 10⁻⁴ →1 × 10⁻⁶; iterations = 10⁵

Feature channels = 96; RRDB blocks = 23; upscale factor = 4; learning rate = 1 × 10⁻⁴→1 × 10⁻⁶; iterations = 2 × 10⁵

Optimizer/Learning Strategy

–

Adam (β₁=0, β₂=0.9)

Adam (β₁=0.9, β₂=0.999)

Loss Function

–

Adversarial loss + L1 reconstruction loss + consistency loss

Perceptual loss + adversarial loss + L1 reconstruction loss

Highlights & Improvements

Preprocessing for stable edge extraction and clean input generation.

Incorporates improved Canny edges as priors to enhance edge consistency and completion accuracy.

Removes BN layers to avoid gradient vanishing; enhanced U-Net discriminator improves edge sharpness and texture realism.