Discriminative Shape Feature Pooling in Deep Convolutional Networks for Visual Classification

Abstract

Unlike conventional handcrafted feature extractors, deep learning approach can extract generic image features without relying on explicit domain knowledge. Recently, there is a trend of combining handcrafted features with learned deep networks to leverage benefits of both. However, the usage of handcrafted features in existing methods are either by naïve concatenation or brute force from deep networks, and lack in addressing issues of parameter quality in the network. In this research, we propose a method that enriches the deep network features by injecting perceptual shape features - Generic Edge Tokens and Curve Partitioning Points, to adjust network’s internal parameter updating process. Thus, the modified convolutional neural network (CNN) produces image representation that tightly embraces benefits from both handcrafted and deep learned features. Our experiments on several benchmark datasets show improved performance compared to models using either handcrafted features or deep network representations alone, with reduced computation and faster convergence rate.