Bottom-up approach to knowledge extraction from digital images

With the rapid advancement of computer vision and image processing technologies, digital images now contain increasingly rich information. However, extracting meaningful high-level semantic concepts from low-level visual features remains one of the core challenges in image analysis. Traditional approaches predominantly focus on the extraction and analysis of low-level features such as colour, greyscale, and edges. Yet these methods often struggle to effectively establish connections between low-level features and high-level semantic concepts. To address the challenge, this paper proposes a bottom-up image knowledge extraction method. Starting from each pixel, it converts the pixel’s RGB values into the Lab colour space and employs a two-dimensional Gaussian distribution model to directly infer each pixel’s high-level semantic concept (such as colour, greyscale value, and whether it is an edge point). Experiments conducted using colour classification demonstrate that this method effectively outputs pixel colours while exhibiting strong interpretability and universality. This approach provides an efficient, intuitive knowledge extraction framework for image understanding, object recognition, and intelligent vision system development, offering both theoretical underpinnings and technical pathways for bridging the gap from perception to cognition.