2026· Poster Volume 0007 The 2026 Twenty-Second International Conference on Intelligent Computing July 23-26, 2026 Toronto, Canada· pp. 3973-3984· 0 citations
TL;DR
This work proposes SLRNet, Super Lightweight Residual Network, a high efficient-yet-effective end-to-end dehazing architecture that integrates a novel Adaptive Feature Unit that automatically adjusts channel-wise features through a lightweight gating mechanism, coupled with compact residual blocks to preserve critical structural information.
Abstract
Image dehazing aims to generate the haze-free images from the hazy observation images. While recent deep learning approaches achieve impressive restoration quality, they suffer from excessive computational complexity and model size, hindering practical applications for real-world deployment on resource-constrained edge devices. To address the limitation, lightweight models are proposed to this end but compromise on dehazing performance. To bridge this gap, we propose SLRNet, Super Lightweight Residual Network, a high efficient-yet-effective end-to-end dehazing architecture. SLRNet integrates a novel Adaptive Feature Unit that automatically adjusts channel-wise features through a lightweight gating mechanism, coupled with compact residual blocks to preserve critical structural information. Unlike standard channel attention mechanisms that discard spatial information, our AFU employs an asymmetric split strategy to simultaneously preserve local texture details and capture global haze density. Our design emphasizes minimal parameter count and low latency without sacrificing perceptual quality. Experiments are carried out across standard benchmarks, showing that our proposed SLRNet demonstrates remarkable performance by achieving state-of-the-art efficiency-accuracy trade-offs compared to existing works, while maintaining robust generalization to real-world haze despite the synthetic-to-real domain gap. The codes are released in https://anonymous.4open.science/r/SLRNet.
A dehazing framework named DKS-Net is proposed which fully utilizes the physics guiding features and extracting structural information in the spatial domain, and a Kernel Selective Feature Extraction Module (KSFE) is introduced to effectively captures structural patterns via large-kernel convolutions with dynamic selection capabilities and multi-scale semantic cues.
Zehao Shi, Han Wang, Xinyue Liu· International Conference on...· 0 citations
Image stitching aims to construct wide field-of-view scenes from multiple narrow-FoV images, yet existing deep learning-based approaches may introduce substantial computational and parameter overhead, limiting their applicability in efficiency-sensitive scenarios. To address this issue, we propose a lightweight deep stitching framework that integrates multi-scale feature fusion with attention-enhanced matching. Specifically, a transformer-based channel attention (TCA) block improves the discriminative capability of fused features in low-texture regions and enhances global consistency. A coordinate-aware correlation module (CACM) combines correlation-based matching with position-sensitive coordinate attention to support registration under parallax, while GhostNet serves as the shared backbone. On UDIS-D, the complete model achieves a 25.19 dB peak signal-to-noise ratio (PSNR) and a structural similarity index (SSIM) of 0.833 under the overlap-region protocol, with a reported full-system complexity of 19.28 giga multiply-accumulate operations (GMACs) and 52.86 M parameters. These results demonstrate a favorable accuracy–efficiency trade-off under the stated GMAC, runtime, and memory protocol and the potential of the proposed framework for resource-conscious image stitching applications.
Remote sensing (RS) image dehazing seeks to eliminate complex and non-uniform haze interference, enabling high-quality image restoration. However, existing methods often face difficulties in balancing adaptability to local haze variations and spatial heterogeneity with computational efficiency. To address these challenges, we propose a novel dual-stream perception interaction network, termed DSPI-Net, which dynamically integrates low-level spatial features and multi-scale features to efficiently restore realistic RS images. The proposed network adopts a lightweight local feature extraction module to efficiently capture shallow spatial details and texture information. Meanwhile, a high-resolution hybrid attention Transformer block utilizes a bidirectional interaction mechanism to capture local details and global contextual information in parallel, producing multi-scale features. To address the intrinsic complexities of hazy RS images, a gated group fusion mechanism adaptively fuses local and global features through an adaptive gating mechanism, achieving optimal feature integration and effectively addressing representation challenges posed by local haze variations and spatial irregularities. Extensive experiments demonstrate that DSPI-Net achieves an average peak signal-to-noise ratio improvement of approximately 0.34 dB over the respective second-best methods across four RS dehazing benchmark datasets, while containing only 0.37 M parameters, corresponding to approximately 24.0% of the parameter count of PCSformer-S. Its lightweight design ensures suitability for implementation on resource-constrained edge devices, meeting real-time processing requirements.
Pan-Pan Liu, Ru-Meng Liu, Bai-Jing Liu et al.· Engineering Research Express· 0 citations
Single image super-resolution (SISR) has achieved remarkable progress with deep neural networks, but the ever-growing model depth brings a heavy computational burden that limits practical deployment. Reducing the inference cost of SR networks has therefore become a central concern. A promising direction is to exploit the fact that image regions differ in restoration difficulty and to allocate computation accordingly. Existing content-adaptive methods have shown the potential of this idea, yet they typically rely on maintaining several separate sub-networks, so that the overall computational and storage cost remains high. In this paper, we propose DADE, a difficulty-aware distillation-enhanced network for efficient super-resolution. A single backbone is equipped with multiple intermediate experts corresponding to shallow, medium and deep computational paths that share their parameters, and a lightweight difficulty-aware classifier is jointly trained with the backbone to route each image patch to a suitable expert. To strengthen the shallow paths, we further introduce a knowledge-distillation scheme in which the deepest expert acts as a teacher and supervises the earlier experts during training. Extensive experiments on four standard benchmarks show that the proposed method substantially reduces the computational cost of super-resolution while maintaining reconstruction quality, with only negligible degradation in PSNR and SSIM, and at the same time greatly lowers the number of stored parameters.
Yuxuan Lin· Advances in Engineering Inno...· 0 citations
Mobile image denoising requires both good restoration quality and low computational cost. In addition, it's annoying to collect large-scale LQ-GT clean pairs. As a result, we propose LiteKD-Net, a lightweight knowledge-distilled network for mobile image denoising. First, a physics-guided noise simulation pipeline generates paired training data by adding pixel crosstalk compared with pipelines applied to cameras. Next, we adapt the Real-ESRGAN to identity-resolution denoising and construct a lightweight Student using Lite-RRDB blocks based on depthwise separable convolutions. Third, feature-level knowledge distillation is applied to transfer the Teacher's restoration capability to the Student without introducing additional inference cost. Experiments on real-world datasets show that our model reaches great reduction in runtime and increase in the inference rate with good restoration quality. Our model also reaches the best in all metrics compared with SwinIR. These results indicate that LiteKD-Net provides a great trade-off between restoration quality and computational efficiency.
This paper proposes BinRVR, a binarized RAW video restoration framework that reduces computation and parameters by approximately 96% while incurring only about 4% performance degradation, and develops a Distribution-Aware Binarized Convolution (DAB-Conv) that leverages the statistics of full-precision activations to mitigate quantization errors.
Tianyu Zhu, Ying Fu, Hesong Li et al.· IEEE Transactions on Pattern...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
