Mining Mid-level Visual Patterns with Deep CNN Activations

TL;DR

This paper introduces a method to discover mid-level visual patterns using CNN activations combined with pattern mining techniques, leading to effective image representations for classification tasks.

Contribution

It proposes integrating CNN activations with association rule mining to efficiently find semantically meaningful visual patterns and mid-level elements.

Findings

Outperforms recent CNN-based methods in scene and object classification

Effective discovery of semantically consistent visual patterns

Two novel image representation methods based on patterns and elements

Abstract

The purpose of mid-level visual element discovery is to find clusters of image patches that are both representative and discriminative. Here we study this problem from the prospective of pattern mining while relying on the recently popularized Convolutional Neural Networks (CNNs). We observe that a fully-connected CNN activation extracted from an image patch typically possesses two appealing properties that enable its seamless integration with pattern mining techniques. The marriage between CNN activations and association rule mining, a well-known pattern mining technique in the literature, leads to fast and effective discovery of representative and discriminative patterns from a huge number of image patches. When we retrieve and visualize image patches with the same pattern, surprisingly, they are not only visually similar but also semantically consistent, and thus give rise to a mid-level visual element in our work. Given the patterns and retrieved mid-level visual elements, we propose two methods to generate image feature representations for each. The first method is to use the patterns as codewords in a dictionary, similar to the Bag-of-Visual-Words model, we compute a Bag-of-Patterns representation. The second one relies on the retrieved mid-level visual elements to construct a Bag-of-Elements representation. We evaluate the two encoding methods on scene and object classification tasks, and demonstrate that our approach outperforms or matches recent works using CNN activations for these tasks.