Attention-Aware Laparoscopic Image Desmoking Network with Lightness Embedding and Hybrid Guided Embedding

TL;DR

This work tackles the problem of surgical smoke degrading laparoscopic image quality by proposing L-SAHGNet, a two-stage framework that first estimates smoke distribution with a Smoke Attention Estimator and then reconstructs a desmoked image through Hybrid Guided Embedding guided by both the original image and a predicted smoke mask. The design integrates a Lightness Embedding Module to inject channel-aware priors and a Restricted Field Transformation to stabilize learning of weight terms, with an explicit reconstruction equation $J(x)= I(x) + K(x)I(x) + B(x)I_s(x) + \varepsilon$, where $K(x)=\frac{1}{t(x)}-1$ and $B(x)=-\frac{1}{t(x)}$. Quantitative results show a $2.79\%$ PSNR improvement over state-of-the-art methods and a $38.2\%$ reduction in runtime, achieving real-time performance on real MIS data, while ablation studies confirm the contributions of LE, SPE, and RFT. The approach offers a transparent, ASM-consistent framework that balances smoke removal quality with computational efficiency, making it well-suited for real-time surgical visualization and downstream image analysis. The work provides practical impact by enabling clearer intraoperative views and robust downstream processing in MIS settings.

Abstract

This paper presents a novel method of smoke removal from the laparoscopic images. Due to the heterogeneous nature of surgical smoke, a two-stage network is proposed to estimate the smoke distribution and reconstruct a clear, smoke-free surgical scene. The utilization of the lightness channel plays a pivotal role in providing vital information pertaining to smoke density. The reconstruction of smoke-free image is guided by a hybrid embedding, which combines the estimated smoke mask with the initial image. Experimental results demonstrate that the proposed method boasts a Peak Signal to Noise Ratio that is $2.79\%$ higher than the state-of-the-art methods, while also exhibits a remarkable $38.2\%$ reduction in run-time. Overall, the proposed method offers comparable or even superior performance in terms of both smoke removal quality and computational efficiency when compared to existing state-of-the-art methods. This work will be publicly available on http://homepage.hit.edu.cn/wpgao

Figures (3)

• Figure 1: Visualization of the L-SAHGNet. The SAE predicts the smoke attention map from the input image $I$. The HGE takes the smoke attention map and $I$ as inputs, and predicts a smoke mask $\tilde{I}_s$ based on the ASM. Additionally, HGE produces a desmoked image $\tilde{J}$ with the guidance of both $\tilde{I}_s$ and $I$. $I_L$ and $\tilde{J}_L$ represent the initial lightness and the predicted lightness, respectively. Lightness is the L of HSL color model. The SPE is an up-sampling module. The RFT block transforms the smoke attention map into the weight matrices.
• Figure 2: Structure of the lightness embedding module.
• Figure 3: Comparison of the proposed method with the state-of-the-art methods on synthetic dataset and real dataset. $I_s$ is the ground truth of smoke mask.