计算机科学
人工智能
计算机视觉
图像(数学)
图像处理
卷积神经网络
图像分割
图像压缩
模式识别(心理学)
算法设计
图像复原
迭代重建
特征提取
卷积码
可视化
目标检测
变换编码
图像质量
作者
Guangfa Wang,Xiang Chen,Jinshan Pan,Jian Jun Yang,Jun Li
标识
DOI:10.1109/tcsvt.2026.3742625
摘要
Image dehazing is crucial for enhancing the robustness of high-level vision tasks in adverse weather conditions. While recent Transformer-based models deliver impressive performance by capturing long-range dependencies, their computational complexity increases quadratically with spatial resolution. At the same time, efficient CNN-based dehazing networks remain constrained by spectral bias and limited receptive fields. To overcome these limitations, we propose DehazeNeXt, an asymmetric encoder–decoder CNN that incorporates three lightweight modules to achieve a superior balance between dehazing quality and computational efficiency. Specifically, we introduce a Generalized FFT-based Frequency Loss (GFFT-Loss), which combines continuous and band-wise perspectives to mitigate the spectral bias inherent in CNNs. We then develop a simple yet effective Dual-input Feature Seesaw Module (DFSM) to adaptively balance multi-level features, promoting the propagation of useful information in later stages. Finally, to expand the limited receptive fields of convolutional operators, we optimize a Dynamic Local Refinement Module (DLRM) for enhanced local feature aggregation. Experiments on synthetic and real-world datasets demonstrate that our method achieves a better efficiency-performance trade-off compared to state-of-the-art approaches. The source code of our DehazeNeXt is available at https://github.com/Bruce-WangGF/DehazeNeXt.
科研通智能强力驱动
Strongly Powered by AbleSci AI