Although historical aerial photographs provide crucial information for efficient long-term environmental monitoring and change detection, they have not been fully utilized due to various formats and image quality factors. At the same time, it has been proved through experiments that knowing spectral characteristics such as color, texture, and brightness can help improve image analysis. It would be a great contribution to colorizing grayscale images!
Traditional Techniques
Until the mid-2010s, various approaches by computer-based colorization were mainly in a semiautomatic manner. These methods were either based on color transfer or color scribbles.
Color transfer is to give the target photo a reference color photo, then the computer will absorb similar spectral characteristics such as color, light, and texture, and then apply it to the grayscale photo to be colored. On the other hand, color scribble is painting colors directly on each area of the target photo in a graffiti manner, and the computer will judge the range and shade changes of the images for coloring.
However, both methods require the participation of manual operations. For example, color transfer needs to provide a reference image, and color scribble needs to manually animate the colored position, which is unsuitable for processing historical aerial images with a large amount of data.
Deep Learning-Based Techniques
Fortunately, the development of deep learning techniques provides insight. First, someone proposed using the Convolutional Neural Network (CNN) method. More than half of the current colorization technology also uses CNN technology. However, regression loss functions, mostly based on Euclidean distance, produce blurry and unsaturated outputs as they tend to minimize the prediction error, leading to low saturation or even ambiguous results which are also not what we want in some cases.
GAN
We use a generative adversarial network (GAN) method for automatic colorization. GAN is composed of two networks, a generator network and a discriminator network, which produce results closer and closer to the target by confronting each other. In addition, compared with CNN needs to provide a large amount of training data to AI for training, this method can reduce the amount of training data and achieve the purpose of automation at the same time.
CIELAB
The common way to describe the color in digital is RGB, that is, the color of each pixel is described by the values of the three primary colors of light: red, green, and blue, which means that we need to predict three values to colorize the grayscale images. Therefore, we use the CIELAB color space, which is defined by the International Commission on Illumination, to generate colors. It is a 3D color space with x and y axes (a*, b*) representing hue and z-axis (L*) representing lightness. Because grayscale photos already have lightness (L*) information, so we only need to predict two values to generate the color, improving the accuracy.
Results of our research and development
We constantly try different settings and data in training, hoping to find the training model that best meets the needs of aerial surveys. After several tests and training, we have gradually trained a more suitable model. The following are the results :
This technique can significantly increase the value of old data by colorizing grayscale images as if giving it a soul. Applied to aerial surveys, it is not only easier to distinguish the contents of the images, such as which side is a forest, which side is a city, etc., but also can be used as important information for exploring environmental changes and spatial and temporal development, which will bring great benefits to mankind!
For image processing and data transformation, DataXquad meets all your needs!
DataXquad is a Pay-Per-Use online image data transformation service platform that can meet all procedures without use costs. We simplify the most complicated part of image data transformation so that more image data can be utilized and more industry chains can derive value from it.
References：
Comentários