Image enhancement is essential in the field of image processing since this improves image quality by highlighting useful information and reducing redundant information in the image.

Image enhancement is among the most essential technologies in the world of image processing, to improve image quality for specific applications. In general, the basic premise of image enhancement is to adjust an image's information contribution such that it is more appropriate for a certain application.

Traditional image enhancement techniques rely heavily on spatial and frequency domain processing. The spatial image enhancement [1] method involves directly processing the pixels in the image, such as the classic modified histogram methods [2–4] and the improved unsharp mask methods [5–7].

The frequency domain image enhancement process involves switching an image to the frequency domain using a mathematical function such as Fourier transform (FT), discrete cosine transform (DCT), or discrete wavelet transform (DWT), then operating image processing based on the frequency domain's unique properties, and finally converting it back to the original image space.

The key benefits of spatial domain image enhancement are its simplicity, minimal complexity, and real-time application.

However, the spatial domain image enhancement method has serious downsides, such as a lack of adequate stability and imperceptibility criteria [8].

It is challenging to propose a solution that can enhance all images. This is mostly due to the following factors: the non-university of the image enhancement process, the selection of the evaluation index, the influence of noise, the selection of optimal parameters, and others.

In this study, image enhancement methods were reviewed in two aspects: supervised and unsupervised algorithms.

Unsupervised Methods

Unsupervised methods, such as K-means [9], hierarchical clustering [18], and the EM algorithm [10], do not require training samples or labels and instead directly model the data.

In the topic of image improvement, we examine various classic unsupervised approaches, including histogram specification, the retinex model, and the visual cortex neural network.

2.1 Modifying the Histogram of Image

Histogram specification [11] is an image processing method that uses image histograms to alter contrast. The brightness on the histogram can be better scattered this way. This can be used to boost local contrast while maintaining overall contrast.