This paper presents a method of compressing and reconstructing a real image using its feature map and a feature catalogue that conprises of feature templates representing the local forms of features found in a number of natural images. Unlike most context-texture based techniques that assume all feature profiles at feature points to be some form of graded steps, this method is able to restore the shading in the neighbourhood of a feature point close to its original values, whilst maintaining high compression ratios of around 20∶1.