Investigating and Explaining the Frequency Bias in Image Classification

TL;DR

This paper explores the frequency bias in CNN image classification, revealing that models favor low- and mid-frequency components over high-frequency ones, influenced by dataset spectral density and class consistency.

Contribution

It provides new insights into the spectral factors affecting frequency bias and investigates their impact on feature discrimination and learning priority in CNNs.

Findings

Spectral density influences learning priority.

Class consistency affects feature discrimination.

High-frequency components are underutilized in CNNs.

Abstract

CNNs exhibit many behaviors different from humans, one of which is the capability of employing high-frequency components. This paper discusses the frequency bias phenomenon in image classification tasks: the high-frequency components are actually much less exploited than the low- and mid-frequency components. We first investigate the frequency bias phenomenon by presenting two observations on feature discrimination and learning priority. Furthermore, we hypothesize that (i) the spectral density, (ii) class consistency directly affect the frequency bias. Specifically, our investigations verify that the spectral density of datasets mainly affects the learning priority, while the class consistency mainly affects the feature discrimination.