A comparative analysis of binary patterns with discrete cosine transform for gender classification

This paper presents a comparative analysis of binary patters for gender classification with a novel method of feature transformation for improved accuracy rates. The main requirements of our application are speed and accuracy. We investigate a combination of local binary patterns (LBP), Census Transform (CT) and Modified Census Transform (MCT) applied over the full, top and bottom halves of the face. Gender classification is performed using support vector machines (SVM). A main focus of the investigation is to determine whether or not a 1D discrete cosine transform (DCT) applied directly to the grey level histograms would improve accuracy. We used a public database of faces and run face and eye detection algorithms allowing automatic cropping and normalisation of the images. A set of 120 tests over the entire database demonstrate that the proposed 1D discrete cosine transform improves accuracy in all test cases with small standard deviations. It is shown that using basic versions of the algorithms, LBP is marginally superior to both CT and MCT and agrees with results in the literature for higher accuracy on male subjects. However, a significant result of our investigation is that, by applying a 1D-DCT this bias is removed and an equivalent error rate is achieved for both genders. Furthermore, it is demonstrated that DCT improves overall accuracy and renders CT a superior performance compared to LBP in all cases considered.