Research On Image Bit-Depth Enhancement Based On Deep Learning

With the development of hardware technology,increasing number of monitors offer displaying high bit-depth images to provide enjoyable visualization with finer color graduations,smoother gradients and more details,which can improve the visual quality significantly.However,most mainstream images and existing digital image acquisition equipment are of lower bit-depth.Besides,some image loss compress operations can also decrease the content bit-depth.These images suffer from false contour artifacts when they are linearly de-quantized since most colors are lost.Therefore,reconstructing visually pleasant images is of vital importance.In this paper,a deep learning based image bit-depth enhancement algorithm is proposed,the main contributions of which are listed as follows:1)We propose an image bit-depth enhancement algorithm based on deep neural network.Skip connections are introduced to enhance gradients of bottom layers in back-propagation,which can stable the gradients and help with training the network.The model reconstructs high bit-depth images directly with suppressed false contours,and the average PSNR results are about 0.4dB higher than traditional ones.2)To enhance feature fusion,we proposed a neural network which concatenates all level features.The proposed network can pass more structural features to top layers without burying their specificity and enhance the vanishing gradients in back-propagation.Besides,we propose to reconstruct residual images rather than reconstructing high bit-depth images directly,leading to better image reconstruction visual quality as well as about 3.9dB higher PSNR results.3)Loss function is of vital importance for deep learning algorithms,which can hardly evaluate image quality thoroughly.Therefore,to better assess reconstructed image quality,we proposed to apply discriminator network to learn to distinguish fake images from real ones.The discriminator network is trained adversarially with the generator and helps with increasing image recovery quality.The average PSNR results are about 4.5dB higher than traditional algorithms.