Remote Sensing Image Colorization Based on Deep Neural Networks with Multi-Scale Residual Receptive Filed

Abstract

To solve the problems of mistaken coloring and color bleeding in the current colorization methods, an end-to-end deep neural network is proposed to achieve remote sensing image colorization. First, the multi-scale residual receptive filed net is introduced to extract the key features of source image. Second, a color information recovery network is con-structed by using U-Net, complex residual structure, attention mechanism, sequeeze-and-excitation and pixel-shuffle blocks to obtain color result. NWPU-RESISC45 dataset is chosen for model training and validation. Compared with other color methods, the PSNR value of the proposed method is increased by 6-10 dB on average and the SSIM value is increased by 0.05-0.11. In addition, the proposed method also achieves satisfactory color results on RSSCN7 and AID datasets.

Remote Sensing Colorization Based on Bidirectional Macro-Micro Adaptive Enhancement Network

“…Feng et. al [34] proposed the multi-scale residual receptive field net to extract the essential information of the original color image and then used U-net with an attention mechanism to recover the color. Subsequently, Feng et.…”

“…f: Feng et al's Method (2021) [34] Use multiple multi-scale residual receptive filed blocks (MR-RFB) to extract features, and then use U-Net as the basic structure to build a color information recovery network (CIR-Net) to recover color information.…”