Author(s): Dang Phan Thu Huong¹, Doan Thanh Binh²

Abstract: Low-light remote sensing images often suffer from severe noise and poor contrast due to insufficient illumination and atmospheric interference, making them challenging to analyze in practical applications. To address these issues, this paper proposes a novel Sparse Denoising Convolutional Neural Network (SD-CNN) that integrates the multi-scale, multi-directional properties of the Curvelet transform with the powerful feature-learning capacity of convolutional networks. Specifically, the Curvelet transform is employed as a sparse representation tool to decompose the input image into different scales and orientations, enabling effective separation of signal and noise components while preserving edge and curve structures. The CNN is then trained on this sparse representation to adaptively suppress noise, enhance contrast, and restore fine details that are often lost in traditional image-processing approaches.

Extensive experiments conducted on simulated and real-world low-light remote sensing datasets demonstrate that the proposed SD-CNN significantly outperforms existing baseline methods, including traditional wavelet-thresholding techniques and pure CNN-based models. Quantitative metrics such as PSNR, SSIM, and visual assessments consistently show that SD-CNN yields higher reconstruction quality and better edge preservation, especially under strong Gaussian noise conditions (σ = 25). Moreover, this hybrid architecture reduces the number of model parameters by approximately 30% through sparse processing in the Curvelet domain, resulting in faster training and inference. Additionally, a dynamic batch normalization mechanism is introduced to enhance training stability and improve model convergence.

In summary, the proposed SD-CNN not only provides superior denoising and enhancement capabilities for low-light remote sensing images but also offers a parameter-efficient design that is well-suited for practical deployment. The findings highlight the potential of combining mathematical transforms with deep learning to tackle challenging image restoration problems in remote sensing and other computer vision tasks.

Keywords: Low-light remote sensing images; Curvelet-based sparse representation; Deep convolutional neural networks; Image denoising and enhancement; Multi-scale and multi-directional image analysis; Dynamic batch normalization; Noise-robust image restoration; Edge and structure preservation.

DOI:10.61165/sk.publisher.v12i6.5

Download Full Article from below:

Curvelet–Deep Learning Fusion Approach for Denoising and Enhancing Low-Light Remote Sensing Images

Pages:41-48