Locally Aggregated Hierarchical Decomposition Based Ensemble Learning for Robust Face Recognition

Abstract

For the detection and extraction of features, a number of hierarchical layering methods have been explored extensively. This is followed by content-based image indexing and retrieval implemented by taking local aggregation of the extracted features. Through a clustering process, a mapping between the numerous patches in the image to the centres of the learned clusters is made. A histogram representation reflecting each of the independent features together form the vectorized encoded image. A global feature vector is then obtained. A comparative study is made with the traditional feature extraction and aggregation techniques following which we extensively explore various learning algorithms to enhance the stability and performance. In preference to using the existing hypothesis space from which the learning models are derived from, we make use of a hybridized hypothesis space which is a combination of the existing hypothesis spaces to get an enhanced hypothesis set. The key idea behind fusing multiple models is to decrease bias which arises from the spurious assumptions that might be concluded by the model, decrease the variance owing to the sensitivity due to minor aberrations that creep in the training phase and to improve predictions. We extend this concept of ensemble system by combining conceptually dissimilar base classifiers and use the majority vote for final prediction. This is performed to balance out limitations of the individual classifiers. A five-fold cross-validation technique was employed on the standard datasets like Grimace and Faces96. Our proposed model can achieve the best accuracy of 99.7 percent on the Grimace dataset and constantly hit in excess of 95 percent for far more complex datasets which are a direct implication of the high tolerance to variances in scale, rotation, pose and expression.