Image classification is one of the most important and much explored problem in machine learning. Evolutionary computation techniques have been applied for image classification for a lesser degree. Even where it is applied to images, it is almost always applied to features pre-extracted from images and not directly at the pixel level. This makes the decision about the kind of features to be extracted a manual one. This paper presents a novel approach for image classification using Learning Classifier Systems that does not require manually designed features. This new technique known as RXCSi works directly on the image pixels and evolves features required for given classification problem. Experiments are performed on MNIST dataset and the initial results validate the effectiveness of the technique for image classification. Another contribution of the paper is that we demonstrate how the evolved features can be visualized to understand basis for making a decision. We believe that RXCSi will open a new direction of research for image classification using evolutionary techniques.