Read the original document by opening this link in a new tab.
Table of Contents
Abstract
Introduction
Related Work
Proposed Approach
Problem Formulation for Cost-Sensitive Classification
Our Proposed Cost Matrix
Properties of the Proposed Cost Matrix
Text
Summary
This paper discusses the issue of class imbalance in real-world object detection and classification tasks. It proposes a cost-sensitive deep neural network approach to address this problem by automatically learning robust feature representations for both majority and minority classes. The paper demonstrates the effectiveness of the proposed approach through experiments on major image classification datasets, showing superior performance compared to baseline algorithms. The key contribution includes introducing cost-sensitive versions of widely used loss functions, analyzing the impact on backpropagation, and proposing an algorithm for joint optimization of network parameters and class-sensitive costs. The paper concludes with the results of extensive testing on various datasets, highlighting the improved performance of the proposed method.