Relationship between feature engineering approaches and deep neural networks on large-scale datasets

Abstract

This paper introduces the relationship between traditional feature engineering approaches and recent deep neural network approaches on image recognition tasks. The tasks aim to understand objects and their situations from given images as we can understand flexibly. Feature engineering approaches are based on statistical modeling in natural language processing and enable to process a large-scale image collection. In general, these approaches first describe local image features from a lot of image patches, then the local features are encoded into a global image feature. Here, the encoding phase purposes to represent compact feature vectors while retaining the information. Both describing local image features and encoding them into a global image feature are developed from empirical and theoretical knowledge of researchers and engineers. Recently, deep neural networks have dramatically improved accuracy or error in several tasks and the knowledge has been applied to extend the network models. This paper briefly introduces the relationship between the feature engineering approaches and the recent extended neural network models.