Liu J, Liu W, Sun J, Zeng T (2021) Rank-one prior: toward real-time scene recovery. In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp 1674–1682 ACM Trans Graph 34(1):1–14īerman D, Avidan S, et al (2016) Non-local image dehazing. IEEE Trans Image Process 24(11):3522–3533įattal R (2014) Dehazing using color-lines. Zhu Q, Mai J, Shao L (2015) A fast single image haze removal algorithm using color attenuation prior. IEEE Trans Pattern Anal Mach Intell 33(12):2341–2353 He K, Sun J, Tang X (2010) Single image haze removal using dark channel prior. McCartney EJ (1976) Optics of the atmosphere: scattering by molecules and particles. Lauer J, Zhou M, Ye S, Menegas W, Schneider S, Nath T, Rahman MM, Di Santo V, Soberanes D, Feng G et al (2022) Multi-animal pose estimation, identification and tracking with deeplabcut. Mo Y, Wu Y, Yang X, Liu F, Liao Y (2022) Review the state-of-the-art technologies of semantic segmentation based on deep learning. Kumar A, Srivastava S (2020) Object detection system based on convolution neural networks using single shot multi-box detector. Additionally, the proposed method exhibits the lowest model complexity while achieving optimal performance metrics and the highest FPS, indicating both its superior dehazing performance and low complexity. The PSNR on the SOTS indoor and outdoor test sets reaches 31.73 dB and 29.31 dB, respectively, with a network parameter size of merely 2 M. Experimental results on the RESIDE benchmark dataset demonstrate that, compared to other advanced methods, the proposed approach achieves superior dehazing outcomes for both synthetic and real haze images, effectively mitigating artifacts, distortions, and incomplete dehazing. In this network, depthwise separable convolutions replace traditional convolutions, significantly reducing model complexity while maintaining satisfactory dehazing performance.
Lastly, a loss function is designed by incorporating contrast regularization and edge loss strategies, effectively guiding the network to generate more realistic images. Subsequently, high-frequency and low-frequency features are merged to reconstruct a clear image. In contrast, for high-frequency features, a detail enhancement module based on deformable convolution is designed to restore fine texture information. For the low-frequency features, which are significantly impacted by haze, a pyramid dehazing module based on large-kernel dilated convolutional attention is devised, facilitating efficient dehazing through complementary semantic information. Initially, Attention Context Encoding (ACE) is employed to decompose the input image into high-frequency and low-frequency features. To address the issues of incomplete dehazing and low dehazing efficiency in existing dehazing networks, this study introduces a Lightweight Contrast-Regularized Dilated Attention Network (LCDA-Net) for single-image dehazing.
0 kommentar(er)
