Chip and System Design of Intrinsic Image Decomposition and Enhancement Based on Conditional Generative Adversarial Network

A new Conditional Generative Adversarial Network architecture the paper proposed, which has two layers of neural networks. The first layer neural network uses the concept of intrinsic image decomposition to train a model that can transfer styles and generate images of reflectance layers. The intrinsic image decomposition algorithm generates the shadow layer image. The second layer of the neural network is used to train luminance correction, and the image of the shadow layer is enhanced to achieve the effect of bright and dark correction. Finally, results of two layers are combined to finish the enhancement, we implement the convolutional layer into the digital circuit through the arrangement of the array, and reducing the computation time.