Ensembling Handcarafted Features to Representation Learning for Content Based Image Classification

Content Based Image Classification (CBIC) has garnered massive interest within research community for it’s applications and significances in multiple areas. This paper has explored a robust feature vector definition scheme based on fusion of handcrafted features to neural network based feature extraction, popularly known as representation learning. The extracted handcrafted features are primarily evaluated with classification results to identify the best color space for significant descriptor definition. Further, feature vector extraction using pre-trained Convolution Neural Network is exercised for evaluating classification performance with representation learning features. Finally, an early fusion of handcrafted feature and representation learning feature is accomplished to examine the classification performance of the model. The test bed is prepared using Wang dataset having 10 different categories of 1000 images. Six different color spaces are explored to finalize the best fit for feature extraction with the proposed handcrafted technique. Performance analyses of the proposed model have turned out to be superior compared to state-of-the-art techniques and have established the importance of fusion based approach in enhancing the accuracy of content based image classification.