WiMi Develops Convolutional Neural Network-based Image Enhancement Algorithm

With the continuous progress of digital image imaging device technology, the number of images acquired through cameras, cell phones, video surveillance, etc., has shown an exponential growth trend, and images have become an essential means for people to perceive the world and exchange information with the outside world. Everything the human eye sees can be translated into the form of images. In the process of image acquisition, generation, compression, storage, and transformation, the complex imaging factors in reality (such as noise, low light, camera shake, object motion, etc.) will lead to degradation of image quality (such as noise, blur, distortion, etc.) and reduce the visual perception quality of images. In this context, the recovery and image enhancement of low-quality images have become a hot research topic in academia and enterprises.

Image enhancement is achieved mainly through contrast stretching, preservation, and high restoration of image details. Firstly, for the whole image, the contrast between the image part and the whole should be improved, and the parties should not be neglected; secondly, the signal-to-noise ratio of the image should be improved to suppress the generation of noise and improve the quality of the photo to make it conform to the visual characteristics of human eyes.

At present, image enhancement is one of the popular directions in the field of digital image processing, and the image enhancement algorithm is mainly a series of processing of the image captured by the imaging device to enhance the overall effect of the image or local details, and various image enhancement algorithms are emerging.

It is understood that the R&D team of WiMi Hologram Cloud, Inc. (NASDAQ:WIMI) is studying image enhancement algorithms based on convolutional neural networks. Convolutional neural networks have outstanding achievements in many fields, such as computer vision and natural language processing. Applying convolutional neural networks to image enhancement has obvious advantages and can solve the challenges of image enhancement in different environments.

The essence of the convolutional neural network is to map the input image to a new mathematical model by multiple data transformations or dimensionality reduction. Convolutional neural networks mainly consist of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The convolution layer performs the convolution operation on the input image or the output features of the previous layer, calculates the inner product of the whole convolution kernel and the corresponding position of the input image or feature map, and extracts the relevant image feature map. The role of the pooling layer includes dimensionality reduction of the activation feature map, reducing the number of parameters and computation in the network, maintaining the feature scale invariance property, and reducing overfitting to a certain extent. The pooling layer can downsample the image using the basis related to the image part to reduce the amount of computational data and leave valid information values. After multiple convolutional pooling operations on the image, the convolutional neural network classifies the features through the fully-connected layer by using the one-dimensional activation feature vector obtained after expanding the three-dimensional activation feature map as the input to the fully-connected layer.

WiMi’s convolutional neural network-based image enhancement algorithm has substantial advantages in extracting image feature information and feature representation. The convolutional neural network can share weights and perform convolutional computation, has strong feature learning capability and mapping capability, ensures the suppression of noise and image detail protection, and has high invariance in image displacement, scaling, and other deformation. It also suppresses noise and image detail protection and has high invariance during image shifts, climbing, and other deformations, showing better-reconstructed image quality.

Convolutional neural networks can learn complex hierarchical features of images and perform complex image recognition tasks. At the same time, the feature extraction based on the convolutional neural network can understand the deep semantic feature information of the image, which can well capture the contextual content of the image, repeatedly train and learn the input image, and finally obtain the best image enhancement effect to meet the requirements of a human vision system for the idea.

Image enhancement algorithms based on convolutional neural networks have broad applications in security, medicine, ecological environment, and other fields. With the rapid development of global information technology, the understanding of the world is increasingly dependent on the explosive transmission of information. The primary way for most people to know the world is still the visibility of the eyes; therefore, image is not only a carrier of human visual information but also an essential medium for spreading information. People’s requirements for image quality are increasing to obtain adequate information in images quickly, the demand for image enhancement will also grow, and the application field of image enhancement technology will be further expanded.

In the future, WiMi’s convolutional neural network-based image enhancement algorithm technology will strive for better progress and more significant breakthroughs in visual effects, contrast ratio, signal-to-noise ratio, etc., laying a solid technical foundation for its more important role in more industries.