DDV-GNet: High-Throughput Defect Detection for Space Manufacturing

Official PyTorch Implementation of the paper: "DDV-GNet: High-Throughput Defect Detection for Space Manufacturing via Deep Delta Gated Networks"

TL;DR

Real-time defect detection in aerospace manufacturing requires processing speeds exceeding 600 Frames Per Second (FPS). Traditional models like ResNet-50 and Vision Transformers (ViTs) fail to meet this threshold without sacrificing accuracy.

DDV-GNet is a novel hierarchical gated convolutional architecture that solves this bottleneck. By integrating Deep Delta Operators with linear-complexity feature gating, DDV-GNet achieves 87.91% accuracy at a blistering 853.94 FPS on an NVIDIA T4 GPU, making it the premier choice for synchronized satellite component production lines.

---

Key Features

• Unmatched Speed: Processes images at 853.94 FPS with a latency of just 1.17 ms, comfortably beating the 600 FPS industrial requirement.
• Deep Delta Gated Architecture: Utilizes gated linear transformations and multiplicative delta rules instead of computationally heavy self-attention mechanisms.
• Domain-Specific Transfer Learning: Pre-trained on the EuroSAT satellite imagery dataset, leveraging unique textual patterns (river/highway grids) that map perfectly to solar panel busbars and micro-cracks.
• Lightweight & Edge-Ready: Achieves highly competitive accuracy with only 7.18M parameters ($O(N)$ linear complexity).

---

Architecture

DDV-GNet features a four-stage hierarchical pipeline (64 → 128 → 256 → 512 dimensions) for multi-scale feature extraction. The core DDV Block utilizes a Gated Convolutional Branch paired with a Deep Delta Operator to ensure optimal gradient flow and data-dependent feature filtering.

(Figure 1: Macro pipeline and Micro DDV block architecture featuring Gated Convolutions and Deep Delta Multiplicative rules).

---

Results & Performance

We evaluated DDV-GNet on the PVEL-AD Dataset (12 classes of solar panel defects) against industry-standard efficient architectures.

Performance Comparison (NVIDIA T4 GPU)

Model	Accuracy (%)	Parameters (M)	Throughput (FPS)	Latency (ms)
ResNet-50	90.11	23.53	249.00	4.02
EfficientNet-B0	89.31	4.02	559.00	1.79
ViT-B/16	85.20	85.81	73.00	13.70
DDV-GNet (Ours)	87.91	7.18	853.94	1.17

Deployment Feasibility

DDV-GNet is the only model tested that successfully lands in the "Real-Time Feasible Zone" for industrial aerospace sorting machines.

---

Quick Start & Installation

1. Clone the Repository

git clone [https://github.com/Latchan-Ch/DDV-GNet-Space.git](https://github.com/Latchan-Ch/DDV-GNet-Space.git)
cd DDV-GNet-Space
2. Install Dependencies
3. Download the Dataset
The PVEL-AD dataset is publicly available on Kaggle. Download it and extract it to the data/ directory.
 PVEL-AD Dataset on Kaggle(https://www.kaggle.com/datasets/programmer3/pvel-ad-electroluminescence-pv-defect-dataset)
4. Run Training
5. Evaluate

Citation

If you find this code or our paper useful in your research, please consider citing: Code snippet:

@inproceedings{chhetri2026ddvgnet,
  title={DDV-GNet: High-Throughput Defect Detection for Space Manufacturing via Deep Delta Gated Networks},
  author={Chhetri, Latchan and Kumar, Aman},
  booktitle={2026 IEEE Space, Aerospace and Defence Conference (SPACE)},
  year={2026},
  organization={IEEE}
}