研究目的
To develop a pipeline neural network for enhancing low-light images by combining denoising and enhancement techniques, addressing issues like noise and poor contrast in low-light conditions.
研究成果
The proposed pipeline network effectively enhances low-light images by integrating denoising and enhancement, outperforming existing methods in both synthetic and real-world datasets. It demonstrates the potential of deep learning for image enhancement without manual parameter adjustment, with suggestions for future improvements in dataset size and network complexity.
研究不足
The method may not perform optimally in all real-world scenarios due to subjective evaluation metrics and potential experimental errors. Performance could be improved with a larger dataset, more hidden layers, or better parameter tuning.
1：Experimental Design and Method Selection:
The study uses an end-to-end fully convolutional network inspired by multi-scale retinex and discrete wavelet transformation. It involves designing a pipeline network with denoising and enhancement components, trained using supervised learning with synthetic and public datasets.
2：Sample Selection and Data Sources:
A dataset of 800 normal light images is collected from Google search, UCID, and BSD datasets, with 10 low-light versions generated per image using HSV scaling and gamma transform, resulting in 8,000 pairs. Additional testing uses public datasets like MEF, DICM, and LIME.
3：List of Experimental Equipment and Materials:
Computational resources for training neural networks, software tools like Keras for implementation.
4：Experimental Procedures and Operational Workflow:
Training involves separate phases for denoising-net and LLIE-net, using Adam optimizer with learning rate 10^-3, batch size 64, and 10K iterations. Evaluation is done on synthetic and real-world images using metrics like SSIM, PSNR, ILNIQE, and SNM.
5：Data Analysis Methods:
Quantitative analysis with SSIM, PSNR, ILNIQE, and SNM; qualitative visual comparisons with state-of-the-art methods.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
