Deep ensemble learning method for zero-shot object detection

Document Type

Article

Publication Date

2-1-2026

Abstract

Zero-shot learning (ZSL) is a machine learning paradigm that enables the recognition of object categories for which training instances are available. This is achieved by leveraging semantic knowledge of classes to establish relationships between seen and unseen categories, thereby facilitating the transfer of learned representations (e.g., visual features) across domains. Conventional object detection models are constrained by their reliance on previously encountered concepts, making them less effective for rare or novel categories. In real-world scenarios, many object categories suffer from scarce or absent labeled samples, which poses a significant challenge to robust object detection. To address this limitation, the present study focuses on zero-shot object detection (ZSD), which extends the detection domain by simultaneously localizing and classifying unseen objects based solely on semantic information. Prior works on ZSD have primarily relied on textual descriptions to construct a shared embedding space bridging visual and semantic modalities. However, these studies have not fully harnessed the potential of textual descriptions to enhance detection accuracy. In this paper, we propose a novel zero-shot object detection framework that integrates Random Multimodal Deep Learning (RMDL) for extracting textual features with a Region-based Fully Convolutional Network (R-FCN) for extracting image features. The proposed multimodal fusion enhances the identification of regions containing objects and enables the classification of newly discovered objects. We conducted extensive experiments on three benchmark datasets—Visual Genome, MS COCO, and ILSVRC—to evaluate the effectiveness of the proposed method. Experimental results demonstrate significant improvements in recall and mean average precision (mAP), particularly when textual descriptions are incorporated into the detection pipeline. The findings underscore that combining visual and textual information offers a promising path for advancing zero-shot object detection, providing a robust solution for identifying and classifying unseen objects.

Keywords

Ensemble learning, Multi-modal deep learning, Visual-semantic embedding, Zero-shot object detection

Publication Title

Applied Soft Computing

ISSN

1568-4946

DOI

10.1016/j.asoc.2025.114333

Funders

King Saud University

Volume

187

Publisher

Elsevier Ltd

Share

COinS