Page 59 - Read Online

Zhuang et al. Intell Robot 2024;4(3):276-92 I http://dx.doi.org/10.20517/ir.2024.18 Page 280

Figure 1. The structure of IE-CGAN algorithm. IE-CGAN: Image Enhancement Conditional Generative Adversarial Network.

image features and then performs a linear combination. Crucially, it integrates a skip connection between the
first and third feature maps to preserve fine image details throughout the mapping process. Meanwhile, the
discriminative network distinguishes improved results from labeled images, assisting the generative process
in creating visually striking outputs. Image transformation into compact feature maps precedes applying a
stacked sigmoid function, resulting in normalized probability scores within the range of [0,1]. In addition to
producing high-quality pictures, IE-CGAN can be used to any resolution, achieving excellent results in both
network performance and the range of applications.

Although the images are enhanced, they must be further processed before being fed into the training network.
Generally, RGB images are taken during the daytime, when the image’s background is bright and the target
color is dark. Nevertheless, since IR images are radiometric, the target radiation is generally more substantial,
and the background radiation is weaker, which means that the distribution of light and dark in IR images is
the opposite of that in RGB images. The mainstream target detection algorithms are more suitable for RGB
images than IR images. Thus, the detection accuracy can be improved if the IR images become closer to the
RGB images after pre-processing. In general, images have 256 grayscales. Supposing there is an IR image
whose original grayscale is denoted by 1. After the grayscale inversion process, the grayscale is represented
by 2. Then the relationship between 1 and 2 is expressed as follows.

(1)
2 = 255 − 1

Where 1 and 2 are integers, taking values in [0, 255]. After the above processing, the input images are visually
closer to the RGB images. We name it the IE-CGAN inversion algorithm.

ThecomparisoninFigure2candemonstratethesuperiorityofourmethod. TheIRimagegiveshighercontrast
after the histogram equalization, and the edges of the objects in the picture are more distinct. However, the
consequent problem is more noise points in other positions. The filtering algorithms are not practical for
processing IR images. They do not make the picture clearer and even blur some image details. The IE-CGAN
can significantly improve the contrast of the IR images. Furthermore, the image details and edges are both
enhanced. Our method has the advantages of the IE-CGAN and makes the image closer to RGB grayscale