The intersection of image enhancement, scaling, and metric precision for future AI systems

Digitization and artificial intelligence (AI) are fostering the development of intelligent applications in the domain of computer vision and image processing. The performance of these applications is significantly dependent on image properties (e.g., image quality and image size). It is imperative to study the development of standard benchmarks for accurate, reproducible, consistent, and explainable AI to understand the impact when the benchmarks are within or outside the defined constraints. Moreover, it greatly influences the design of associated tools and applications to understand the performance of the models and distinguish them based on real-world and synthetic data for critical applications (e.g., detection in darkness tasks, etc.). This work presents a comprehensive experimental overview of the classic and learning-based image enhancement methods and also quantifies the impact of the most widely used vision-based methods when serving in the data preprocessing for the learning-based models. The impact of image properties is evaluated through experimental verification using several metrics, which helps to understand the role of image-quality consistency in future AI applications influenced by image properties. This work offers a thorough experimental review to understand the importance of image enhancement (IE) in the context of deployment in smart applications by using standard evaluation metrics and also illustrates the relationship and impact of image scaling on the evaluations. We performed experiments to measure the impact of image properties on image-quality consistency in AI training data and to evaluate performance using evaluation metrics. We also propose the perceptual integrity measure (PIM), a structured, perceptual, feature-based no-reference measure formulated as a deterministic, interpretable scoring function that aggregates meaningful image features into a unified score for quantifying image integrity across diverse imaging conditions. Overall, this work consolidates existing knowledge, provides a comprehensive understanding of this evolving field, and identifies potential directions for future research.