Liquid crystal displays (LCDs) have been widely used in consumer electronics, industrial control, medical equipment and other fields. However, the LC panel can not completely block the light from the backlight unit when displaying a black field, resulting in a low Contrast Ratio (CR). There are two techniques to improve CR, including dynamic dimming and dual-layer display. Dynamic dimming technology dynamically adjusts the backlight brightness and pixel grayscales according to the displayed image. At present, dynamic dimming technology is mainly divided into global dimming technology and local dimming technology. CR is proportional to the number of backlight partitions, but the number of backlight partitions is usually much smaller than the number of pixels on the LC panel, resulting in the inability of this technology to achieve pixel-level dimming. The more partitions, the higher the production cost. Another technology to improve CR is dual-layer display. Dual-layer display is to add a LC panel between the backlight and the LC panel, which can greatly reduce light leakage. However, as the two layer LC panels are bonded together with Optically Clear Adhesive (OCA), there is a physical gap between them. When the double-layer LCD is working, both the front and rear panels need to display images. If the input image is directly sent to the front and rear panels without processing, the corresponding pixels of the front and rear panels will be offset when viewing the screen off-axis, resulting in a ghost image. To improve the display quality when viewing off-axis, the input image needs to be split and modified. The common method is to blur the image to the rear panel near the backlight unit and compensate for the pixels of the front panel. It is quite simple to realize such a blurring solution, but the final display quality deteriorates because the resolution of the image to the rear panel becomes lower. In order to balance the relationship between viewing angle and display quality, researchers put forward an angle compensation algorithm. The algorithm establishes a mapping matrix to record the position information of each ray, and then obtains two images through an optimization algorithm. The algorithm can present a higher display quality at larger viewing angles. But the processing time is quite long and no real application is possible.Inspired by the successful applications of Convolutional Neural Networks (CNNs) in image restoration, the paper proposes to adopt CNN to optimize the image for dual-panel display in order to reduce processing time and improve high display quality. The network structure includes three parts: preprocessing, feature extraction and reconstruction. For a dual-panel display, it is best that the images viewed at various angles are as the same as the image viewed off-axis. Therefore, in the preprocessing part, the input image is duplicated to create N copies corresponding to the N viewing angles before sending the image to the network. The feature extraction part consists of 8 convolution layers, and the shortcut connection is introduced to create the residual block that can fuse the features from the shallow network. In the reconstruction part, a reconstructed image is constructed by corresponding multiplication of two images with pixel offset at a certain angle. The purpose of network training is to minimize the errors between the reconstructed images at different viewing angles and the original input image. Therefore, the mean square error is used as a loss function, which is especially defined with the consideration of the total differences between the reconstructed images at different viewing angles and the input image. For dual-panel display, ghost may appear when viewing off-axis, which presents worse effects on the images consisting of many textures. The paper therefore creates a dataset for a dual-layer display, which consists of many images such as figures, animals, buildings, and scenery. All the images contain rich texture information.In order to evaluate the proposed CNN for dual-layer display, some existing methods are applied for comparisons, including fuzzy processing algorithm and viewing-angle-compensation algorithm. 5 viewing angles of 0°, 8°, 34°, 45° and 64° are used for evaluation. To quantitatively evaluate the three methods, Peak Signal-to-noise Ratio (PSNR) and Structural Similarity (SSIM) are employed, and the computation time is compared. Compared with fuzzy processing algorithm, the proposed method presents obvious advantages in PSNR, SSIM and computation time. The PSNR and SSIM of the proposed method at the viewing angle of 0°, 18°, 34° and 45° are almost the same as the viewing-angle-compensation algorithm. The PSNRand SSIM of the proposed method at the viewing angle of 64° are lower than viewing-angle-compensation algorithm, but much higher than the fuzzing processing algorithm. When comparing the computation time, the proposed method presents a great improvement over the other two methods. The computation time of the proposed method is 707.13 times and 4.75 times that of the other two methods, which presents strong practicability.The paper proposes a CNN-based method for dual-layer display. The CNN is used to process the input image, and residual blocks are constructed to improve the quality of the output image. The two output images that correspond to the two layers of LC panels of the dual-layer LCD are reconstructed from different viewing angles to compare with other methods in terms of computation time and display quality. The simulation results show that the proposed method improves both computation time and display quality compared to the fuzzy processing algorithm; compared with the viewing-angle-compensation algorithm, it can greatly reduce the computation time. The proposed method for dual-layer display presents the merits of high display quality and short computation time, which presents a good application prospect.