Infrared Image Colorization Algorithm Based on DCGAN and Its Edge Device Acceleration

The image collected by the infrared night vision device is a grayscale image, but the human eye can distinguish only dozens of grayscale levels, which restricts the analysis and processing of the measurement image. Infrared image colorization can greatly improve the target tracking and recognition performance, speed and accuracy. This paper proposes an infrared coloring model based on DCGAN and the corresponding model compression algorithm. The generator of the DGCAN model adopts the Unet structure, and the discriminator adopts the PatchGAN structure to ensure better performance in the domain transfer from infrared images to color images. Hinge Loss and L1 loss are used to ensure the stability of model training, while taking into account the edge device transplantation of the model, the network is distilled during the training process, network pruning and neural network structure search and other operations to achieve model compression and deploy on edge devices superior. The experimental results of multiple datasets show that the colorization algorithm proposed in this paper can greatly improve the processing speed while maintaining the colorization performance, and realize the real-time processing of infrared image colorization.