Image enhancement algorithms commonly involve adjusting an image’s brightness, contrast, saturation, hue, and other attributes to improve its clarity and reduce noise. Common methods include histogram equalization, Gamma transformation, Laplace transformation, the Retinex algorithm, and deep-learning-based image enhancement.

Overview

Image enhancement algorithms commonly involve adjusting an image’s brightness, contrast, saturation, hue, and other attributes to improve its clarity and reduce noise.

Image enhancement often involves combining multiple algorithms to achieve the aforementioned functions. For example, image denoising is equivalent to a low-pass filter, while enhancing sharpness corresponds to a high-pass filter. Of course, the primary goal of enhancing an image is ultimately to extract useful information and services from it.

The general algorithmic process can include: image denoising, enhancing sharpness (contrast), grayscale conversion, extraction of image edge features, or convolution-based binarization. These four steps can often be implemented using different approaches. In the following sections, we will conduct specialized experiments on this content, list its application scenarios, and highlight its processing characteristics.

Common Image Classification Algorithms

Histogram equalization

In a histogram, if the gray levels are concentrated in the high-gray-level region, the low-gray-level areas of the image become difficult to distinguish. Conversely, if the gray levels are concentrated in the low-gray-level region, the high-gray-level areas become hard to discern. To ensure that both high and low gray levels are easily distinguishable, the best approach is to transform the image so that the distribution of gray levels becomes uniform. This is precisely the purpose of histogram equalization. Suppose the image undergoes the following transformation:

Here, L represents the number of gray levels. Our goal is to make the probability distribution of gray levels uniform:

The relationship between the gray-level distribution before transformation and the gray-level distribution after transformation is:

Therefore:

Represented using discretization:

The actual effect is shown in the figure below (left: original image; right: result):

Advantages and Disadvantages of the Algorithm: If you directly equalize the R, G, and B channels of a color image separately and then merge them back together, it’s highly likely to result in issues such as uneven colors and distortion. Therefore, it’s common practice to convert the RGB image into the YCbCr color space and then equalize only the Y channel (which represents the luminance component).

Gamma transformation

The gamma transformation is primarily used for image correction, adjusting images with either excessively high or low grayscale levels to enhance contrast. The transformation formula involves multiplying each pixel value:

As shown in the gamma curve in the figure below, you can see its transformation principle:

As shown in the figure, when the value of r is greater than 1, the regions with higher grayscale levels in the image are stretched, while the regions with lower grayscale levels are compressed. Conversely, when the value of r is less than 1, the regions with lower grayscale levels are stretched, and the regions with higher grayscale levels are compressed. This is how image contrast is adjusted.

The actual effects are shown in the figures below (left: r=0.5; right: r=1.5):

Laplace transform

Performing a second-order differential operation on an image can highlight image edges and enhance details. Typically, one aims to construct an isotropic filter that is insensitive to changes in image orientation. The simplest isotropic filter is the Laplace operator, defined as follows:

Add the result of the Laplace transform to the original grayscale image according to a certain proportion:

This can help highlight the contrast.

Retix algorithm

The Retinex theory posits that the world is inherently colorless, and the colors we perceive are the result of the interaction between light and matter. In other words, the image projected onto the human eye is influenced by the long-wave (R), medium-wave (G), and short-wave (B) components of light, as well as the reflective properties of objects.

Here, I represents the image seen by the human eye, R is the reflection component of the object, L is the ambient light illumination component, and (x, y) denotes the corresponding position in the two-dimensional image.

It calculates R by estimating L. Specifically, L can be obtained by convolving the Gaussian blur with I, which can be expressed mathematically as:

Among them, F FF is the Gaussian blur filter, and * denotes convolution operation.

Among them, σ is referred to as the Gaussian Surround Space Constant—the scale used in the algorithm—which has a significant impact on image processing.

The actual effect is shown in the figure below:

Image Enhancement Based on Deep Learning

LL-NET is the first method to enhance images using deep learning. In its paper, the authors address the problem of low-light image enhancement from the perspective of representation learning by employing autoencoders. These autoencoders are trained to learn the underlying signal features in low-light images and adaptively brighten and denoise them. Drawing on the sparsity property of the SSDA network, LL-NET can effectively remove noise from images. By leveraging the network’s generalization capability, LL-NET is applied to test images captured under low-light conditions, enabling the network to learn the intrinsic features of the images, thereby reducing noise and enhancing image contrast.

The network structure is as follows:

The actual effect is shown in the figure below: