Authors: Arpit Urmaliya, Jitendra Kumar Maitra
Abstract: Effective image denoising is a challenging task for real-world smartphone images due to the complex noise resulting from the limited response of the sensors and the varying illumination conditions. The paper proposes and tests a light weight deep learning framework called Convolu-tional Adaptive Multi-Scale Denoising Network CAMSD-Net to restore high quality image from noisy smartphone photos in the meanwhile retaining the structure and visual quality. The pro-posed methodology employs the SIDD-Small sRGB dataset with 160 pairs of noisy and clean images which are systematically validated, with 70% used for training, 15% for validation, and 15% for testing, and with random 256×256 image patches extracted from each image, leading to a fixed 384 paired image patches used for testing. The network is implemented based on multi-scale convolutional network, local spatial learning, dilated contextual feature extraction, feature fusion and skip connections, and is trained for 30 epochs with the Adam optimizer and learning rate of 0.001. Experimental results show the superior performance of CAMSD-Net V2 over classical denoising schemes and the first version of the network. The proposed model achieves the highest Peak Signal-to-Noise ratio (PSNR) of 36.5564 dB, and the Structural Similarity Index (SSIM) of 0.925337, and minimizes the Mean Squared Error (MSE) of 0.00032221, Root Mean Squared Error (RMSE) of 0.016248, and Mean Absolute Error (MAE) of 0.012848. In addition, the proposed framework is effective in real world smartphone image restoration as evidenced by its performance on 356/384 test patches (92.71%) with respect to the improvement of PSNR.
DOI: http://doi.org/10.5281/zenodo.22829914
