A lightweight color depth learning model for near-infrared images with fusion layer

Abstract

The invention relates to an automatic near-infrared image colorization model composed of a lightweight image recognition network module and an image colorization CNN module with a fusion layer. Fast-RCNN performs image recognition on near-infrared image, and then collects and selects images similar to the scene from the network, After training with the colorized CNN module with fusion layer, the near-infrared image is input for colorization. Therefore, the colored near-infrared image is output. This method combines the lightweight image recognition model with the image colorization model withfusion layer, which not only colorizes most of the near infrared images, but also improves the problems such as boundary diffuse color, gray and dark color, coloring error and so on in the colorized images.

Description

Technical field [0001] It relates to a method for automatic colorization of near-infrared images. The method has good colorization effects for low-resolution, high-resolution near-infrared images, and infrared images taken during the day or night, and belongs to image processing technology. It can be applied to the automatic colorization of near-infrared images in the security field. Background technique [0002] Since the near-infrared image (NIR) can reflect the thermal radiation information of the target in the scene, it is not sensitive to the brightness change of the scene, and many important night vision or low-illumination scenes, such as mines, wild animal observation points, military bases, etc. Use near-infrared images to achieve comprehensive monitoring. However, since the near-infrared image obtained is a grayscale image, it also has the shortcomings of blurred edges and insufficient details, so it needs to be colored to increase its color and texture information to ...

AI Technical Summary

Problems solved by technology

First of all, the existing colorization model using deep learning needs to train a very deep network model, which has extremely high requirements on the hardware of the device, and it is difficult to achieve mass production

In addition, the effect of using the deep learning colorization model on image colorization is highly dependent on the image quality and richness of the training set. Most of the existing deep learning colorization models use large data sets such as ImageNet as the training set. Although the training set has a large number of images, However, it also greatly increases the training time of the network, and there is no corresponding image for a certain type of special scene (rare animals and plants, special food, etc.), so the image of this scene cannot be colored

Images