Image recognition techniques have been widely used in positioning systems in recent years. By recognizing the objects targeted by users' camera, one can decide the users' location. In this paper, a mobile indoor positioning system based on the image recognition techniques is implemented for shopping malls. We recognize the stores by their logos, and then use the location of the stores to locate the users. The image recognition method includes extracting local features from the image, calculating the Bag-of-Word structure through a pre-trained hierarchical clustering tree, and using cosine similarity to make the comparison between the training images and the query images. Though SIFT and SURF are the most extensively used local feature detectors and descriptors in the field, the limitations of mobile devices make them infeasible due to their high computational complexity. Moreover, both SIFT and SURF are patent-protected and are not free modules in OpenCV4Android, which will cause additional cost. Therefore, in this paper, we attempt to adopt features that exclude SIFT and SURF. By analyzing the precision and speed of pairwise mashup of feature detectors and descriptors, we target to find the most suitable pair of algorithms to be used on mobile devices. In this paper, the Global Mall at Hsinchu, Taiwan, is used as a scenario for the actual test.