Journal of Pattern Recognition Research

This paper reports the performance of a character recognition test using MNIST handwritten-digit database. The presented assessment is twofold, first it shows the performance of a recognition test based on a very simple feature extraction method, secondly it shows a disparity within the database that may be important when recognition algorithms are compared. For the assessment of the MNIST database, a novel feature extraction scheme based on the Mojette transform is proposed. This transform has not been used as the feature extraction scheme for character recognition. The Mojette transform can be seen as equivalent to the so called projection histograms. Other descriptors, like the Kirsch directional features, require a mask and consequently, many multiplications, so the Mojette transform is mathematically less complex since it requires only sum operations. In spite of the very simple projection method used as a descriptor, a recognition rate of 97% was obtained over a testing set of 1000 samples, as used by some authors. It was found, however that for a fixed-size test of 1000 samples, the recognition rate changed on different sections of the database.