VisionDES

Robust and Explainable Dynamic Vision Ensemble

Overview

VisionDES is a vision-oriented Dynamic Ensemble Selection (DES) framework that adapts classical DES techniques to modern deep learning pipelines.

Instead of relying solely on classifier confidence, VisionDES combines:

• 🎯 Local classifier competence
• 🔍 DINO visual embeddings
• ⚡ FAISS nearest-neighbor retrieval
• 🧠 Feature-space similarity estimation
• 📊 Dynamic competence weighting
• 🔬 Explainable ensemble decisions

For every incoming image, the framework dynamically identifies the most competent classifiers within a local Region of Competence (RoC) and generates an adaptive prediction.

Framework

Installation

pip install vision-des

https://pypi.org/project/vision-des/

Quick Start

Create a Pool of Models

pool = [
    resnet50,
    efficientnet,
    convnext
]

Initialize VisionDES

from vision_des import VisionDES

des = VisionDES(
    dsel_dataset=dsel_dataset,
    pool=pool,
    device="cuda"
)

Build the Retrieval Index

des.fit()

Predict

prediction = des.predict(
    test_image,
    k=7
)

---

Explainable Inference

prediction = des.predict(
    test_image,
    k=7,
    explain=True
)

Sample Output

Example for Attack Scenario

Citation

We would appreciate it if you could cite our work when using our code.

@inproceedings{Juraev_Abuhmed_visiondes_2026,
  author    = {Juraev, Firuz and Abuhamad, Mohammed and El-Sappagh, Shaker and Woo, Simon and Abuhmed, Tamer},
  title     = {VisionDES: Robust and Explainable Dynamic Vision Ensemble},
  booktitle = {Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (KDD ’26)},
  year      = {2026},
  address   = {Jeju, South Korea}
}