WiMi Hologram Cloud Inc. (NASDAQ: WIMI) (“WiMi” or the “Company”), a leading global Hologram Augmented Reality (“AR”) Technology provider, today announced a new and more efficient solution for computer-generated holography (CGH) through deep learning and neural network technology. Deep learning can find the optimal or local optimal solution in operation, making it efficient to compute CGH. CGH has been applied to holographic light traps, 3D displays, planar concentrators, AR displays, etc.
CGH technology can obtain the best wave modulation by inverse-solving the custom light field. Image quality is limited by the accuracy of SLM modulation, which is usually challenging to represent the target light field. In practice, the solution of computational holograms is always approximate and numerical methods are needed to determine feasible holograms to obtain the best-encoded wavefront. The current computation in CGH usually uses iterative algorithms, and non-iterative methods are designed to save computation time by evolving the GS algorithm. Despite the improvement, these non-iterative methods always lead to poor image quality and low spatial resolution during reconstruction due to scatter noise, downsampling effects, and conjugate image interference. In using deep learning technologies, U-net structures have been tried on CGH problems with initial success, but the holograms obtained by U-net in computational holographic problems have the drawback of degrading the quality of reconstructed images. Traditional convolutional neural networks rely on convolutional filters and nonlinear activation functions, which means that the processed data are assumed to be linearly separable. However, problems such as image coding, holographic encryption, and frequency analysis are difficult to describe by linearly divisible functions, and simple convolution and deconvolution are always restricted to a certain region to improve operational efficiency. The inability of U-net to utilize and rewrite global information means that optical image processing is very weak.
WiMi has developed efficient computer-generated holography (ECGH) technology, a deep learning-based CGH imaging method, which aims to solve the problems of long computational cycles and poor quality of traditional CGH methods. The method uses a mixed linear convolutional neural network (MLCNN) for computational holographic imaging and enhances information mining and information exchange by introducing a fully connected layer in the network.
The network uses an MLCNN structure with line fork layers, a “DownSample” structure for down-sampling, and an “UpSample” structure for up-sampling. The technology uses a neural network model to calculate the input target optical field and computes the phase values to simulate the optical experimental results. The target optical field is compared with the simulation results using a loss function, and the gradient of the loss value is calculated and back-propagated to update the network parameters.
WiMi’s ECGH method can quickly obtain the required pure-phase images to generate high-quality holographic imaging. Compared with the traditional deep learning-based CGH method, WiMi’s ECGH technology can reduce the number of parameters required for network training by about 60%, thus improving the efficiency and reliability of the network. In addition, the network structure of ECGH technology is highly versatile and can be used to solve various image reconstruction problems, which has strong practicality and application prospects.
WiMi’s ECGH images use a non-iterative deep learning model MLCNN, which can compute hologram generation faster. By successfully applying the ECGH method, high-quality and stable computational hologram images can be obtained. A major feature of the MLCNN structure is the ability to compute cross-region exchange of data, which makes it suitable for complex optical functions that require the manipulation of global information. Applying the MLCNN model in WiMi’s ECGH technology can effectively handle the complexity of optical functions. The model can handle a variety of complex optical functions to generate high-quality holographic images. This holographic image can perfectly reproduce the 3D scene, giving the observer a more realistic visual experience. The MLCNN model has better optical domain adaptation than the U-net network structure. This gives it an advantage in holographic generation and reconstruction because it can better handle the complexity of optical functions and variations in the optical domain, and CGH can perfectly reproduce the ability of 3D scenes and prevent visual fatigue.
The ECGH technology developed based on the MLCNN model framework of deep learning and neural networks not only reduces the computational load but also improves the quality of holograms, thus making CGH more practical. In addition, the MLCNN model is highly flexible and can be adapted to different holographic generation tasks. It has excellent computational power and high-quality hologram generation capability. With the continuous development of technology, the ECGH technology of the MLCNN model will be more widely used.