Augmenting Deep Learning Performance in an Evidential Multiple Classifier System

Vandoni, Jennifer; Le Hégarat-Mascle, Sylvie; Aldea, Emanuel · 2019 · DOAJ

DOI: 10.3390/s19214664

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of applying deep learning to tasks with limited labeled data, specifically pedestrian detection in high-density crowds. The authors identify two primary issues: the scarcity of precise training data due to occlusion and annotation difficulties, and the "black-box" nature of deep learning models, which lacks interpretable uncertainty measures. To mitigate overfitting and improve interpretability, the study proposes an Evidential Multiple Classifier System (MCS) that fuses deep learning outputs with traditional classifiers within a Belief Function Theory (BFT) framework. The methodology combines two distinct classifier ensembles. The first is a Convolutional Neural Network (CNN) ensemble derived using Monte Carlo dropout. The authors design a specific fully convolutional network, termed FE + LFE, which utilizes dilated convolutions to maintain high resolution for detecting small objects like heads without pooling layers. To handle imprecise ground truth data (dot annotations rather than bounding boxes), they employ a soft-labeling technique using cumulative Gaussian distributions. The second ensemble consists of Support Vector Machine (SVM) classifiers trained on hand-crafted features using an active learning procedure. These two ensembles are treated as independent information sources and fused using BFT, which models both uncertainty and imprecision, allowing for a more robust decision-making process than standard probabilistic fusion. The results demonstrate that this evidential fusion strategy significantly improves detection performance compared to individual classifiers. By explicitly modeling the uncertainty and imprecision inherent in both the deep learning predictions and the noisy training data, the system achieves more reliable pedestrian detection in crowded scenes. The approach allows for a deeper interpretation of results, providing measures of the model's commitment to each decision. The study confirms that deep learning techniques can be effectively applied to small datasets when combined with regularization methods like Monte Carlo dropout and fused with complementary traditional classifiers in an evidential framework. The significance of this work lies in its demonstration that deep learning is viable for specific, data-scarce applications where standard out-of-the-box models fail. It provides a practical method for quantifying model uncertainty without the computational overhead of full Bayesian Neural Networks. Furthermore, the integration of BFT offers a robust mechanism for handling the intrinsic imprecision of real-world annotation data, contributing to more trustworthy and interpretable AI systems in critical domains such as security and autonomous driving.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• distraction detection algorithms