Using Convolutional Neural Networks to Discover Cogntively Validated Features for Gender Classification

• Published in: 2014 International Conference on Soft Computing and Machine Intelligence pp. 33 - 37
• Main Authors: Verma, Ankit, Vig, Lovekesh
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.09.2014
• Subjects: Accuracy; backpropagation algorithm; Biological neural networks; Convolution; convolutional neural network; Face recognition; Feature extraction; gender classification; horizontal flipping; L2 Regularization; Robustness; Training
• DOI: 10.1109/ISCMI.2014.17

The human visual cortex is extremely adept at distinguishing between male and female faces, or performing "Gender Classification". While the subject of face detection and recognition has received a lot of focus, research into the features or cognitive processes that are useful for identifying gender have received relatively little attention. Researchers have attempted to extract hand crafted features like wavelet coefficients, histograms etc. on the basis of which to generate a model to classify the male and female faces. However, these models tend to compress the image into a vector and disregard the two dimensional spatial correlations between the pixels in an image. Additionally, these features have to hand crafted and may or may not be ideal for the classification at hand. Ideally, the system should be able to generate specific features from the input face image which would help in classification of male faces from female faces. In this paper and a Deep Convolution Neural Network (CNN) model is presented for gender classification. The features generated by the CNN appear to agree with known results from the cognitive science community indicating that these models may be closer to biological neuronal processes governing gender classification. The classification results are compared with different regularization techniques and other standard classifiers, and the CNN models yield higher accuracy than both svms and random forest classifiers.