Due to the overwhelming volume of image data present on the Internet and World Wide Web, it is necessary to have new, intelligent methods for image representation and analysis. Semantic-based compression is a technique for conveying high-level contextual information that can describe concepts such as image content, video sequences, 3D environments and sound. This paper describes a method that can obtain compression ratios of up to 2000:1 for high resolution colour natural outdoor images. It will be argued that for fast browsing and retrieval of images within a large database, a cartoon-like image representation, termed a `Quick- Sketch', that allows the user to discern the original image content, is sufficient. Using neural network technologies to obtain a semantic classification of image content, we propose a method that encodes an image as a set of labels and region boundaries. The decoding and reconstruction process then involves the regeneration of region shapes from the boundary information and label based generation of colour and texture for each region using computer graphics techniques.