ECA LDNet A Dual Attention Network with Atmospheric Scattering Guidance for Single Image Dehazing

<title>Abstract</title> <p>Single-image dehazing requires balancing restoration fidelity against model compactness: high-capacity attention and Transformer models achieve strong accuracy, whereas very small convolutional models often sacrifice fidelity. We present ECA-LDNet, a 1.482-million-parameter depthwise separable U-Net that combines reduction-free Efficient Channel Attention (ECA), pixel attention, and a softly blended atmospheric-scattering-model branch. The auxiliary branch predicts latent transmission-like and atmospheric-light-like variables and contributes 8% of the output; these are guidance signals, not physically calibrated estimates. ECA-LDNet achieves the 32.53 dB peak signal-to-noise ratio (PSNR) and 0.9717 structural similarity (SSIM) on SOTS-Indoor, 31.94 dB / 0.9769 on SOTS-Outdoor, and 30.20 dB / 0.9641 on RESIDE-6K. Because training uses a mixed RESIDE-6K and ITS corpus, published cross-method results are treated as protocol-unmatched references rather than controlled rankings. Controlled single-seed ablations show that ECA gives the most consistent PSNR gain per added parameter, whereas pixel attention and the atmospheric branch trade PSNR against structural similarity. At 256 × 256 the model needs 2.171 GMACs by a layer-wise counter released and runs at about 12–14 frames/s on an NVIDIA Tesla P100. Under one identical evaluation pipeline, a 22-image comparison of five released pretrained models gives ECA-LDNet the highest overall PSNR, though the set is too small for population-level ranking, and the models differ in training data. ECA-LDNet is therefore a compact design with explicitly characterized component trade-offs rather than a state-of-the-art result. Code, weights, and evaluation scripts are publicly released.</p>