Image entropy equalization: A novel preprocessing technique for image recognition tasks

Image entropy is the metric used to represent a complexity of an image. This study considers the hypothesis that image entropy differences affect machine learning algorithms' performance. This paper proposes a novel preprocessing technique, image entropy equalization, to delete the image entropy differences. The goal is to transform all images into the same entropy. Such a process is implemented by editing all images into the same histogram. Image entropy equalization is evaluated by comparing the original and equalized images in various machine learning tasks. The main advantage of image entropy equalization is to improve the AUC score for one-class autoencoder (OCAE). This result gives a new hypothesis that using image entropy equalization could improve various studies using autoencoder (AE). In addition, the proposed method shows fair results for classification and regression tasks. On the other hand, the main challenges are that the equalization process depends on a reference histogram and is affected by diverse backgrounds.

Automated summary

This paper proposes a novel preprocessing technique called image entropy equalization to eliminate the differences in image entropy. It compares the original and equalized images in various machine learning tasks, and shows that image entropy equalization can improve the AUC score for one-class autoencoder, as well as achieve fair results for classification and regression tasks.