Faster OreFSDet: A lightweight and effective few-shot object detector for ore images

Yang Zhang, Le Cheng, Yuting Peng, Chengming Xu, Yanwei Fu, Bo Wu, Guodong Sun

Introduction

In the mining industry, the precise measurement of ore particle size is a critical factor for optimizing crushing and grinding processes. While deep learning has significantly advanced the field of object detection, its application in mining environments faces two primary obstacles: the high cost of labeling specialized datasets and the computational intensity of standard detection frameworks.

Faster OreFSDet is a specialized, lightweight few-shot object detector designed to address these challenges. By moving away from heavy traditional architectures, the authors have developed a model that is only 19MB in size and operates at 50 FPS, significantly outperforming existing few-shot methods in both accuracy and deployment efficiency.

The Challenge: Data Scarcity and Deployment Constraints

Traditional object detectors require thousands of labeled images to achieve high accuracy. In the mining sector, obtaining and labeling such a volume of data is both expensive and time-consuming. When standard detectors are trained on only a few samples, they tend to overfit, leading to poor performance on unseen data.

Furthermore, most Few-Shot Object Detection (FSOD) models are built upon the Faster R-CNN framework. While accurate, these models are often too slow and memory-intensive for real-time use on edge devices at beneficiation sites. There is a clear need for a model that can learn from a handful of examples (few-shot) while remaining fast enough for real-world, real-time deployment.

Methodology

The researchers replaced the standard Faster R-CNN backbone with CenterNet2, a more efficient two-stage detector. CenterNet2 utilizes a probabilistic first stage that generates higher-quality proposals with significantly fewer candidates (256 vs. the 1000 typically used in Faster R-CNN). To further enhance performance in the few-shot context, the authors introduced three novel modules:

  • Support Feature Mining (SM) Block: Rather than relying on raw support images, this block uses linear projections to encode features along height and width dimensions. This allows the model to focus on "position-sensitive" information while filtering out irrelevant background noise.
  • Relationship Guidance (RG) Block: This module strengthens the link between the "support" (the example image) and the "query" (the image being searched). By using multi-scale kernels (1x1 and 3x3), it captures both spatial and channel-wise correlations, ensuring that critical support information is preserved during the search process.
  • Dual-scale Semantic Aggregation (DSA) Module: Ore particles naturally occur in various sizes. The DSA module retrieves features at two different resolutions (4x4 and 8x8) during the final prediction stage, ensuring that both small and large particles are detected with high precision.

Experimental Setup and Results

The model underwent "base training" on the MS COCO dataset to establish a general understanding of object features. It was then fine-tuned on a Custom Ore Image Dataset featuring ores of different densities and sizes, using only 5, 15, or 25 labeled samples ("shots").

Performance Benchmarks

Faster OreFSDet sets a new benchmark for industrial few-shot detection across several key metrics:

  • Accuracy: On the 25-shot ore task, the model achieved 54.1 AP, vastly outperforming the baseline AttentionRPN (30.8 AP) and even surpassing general object detectors like YOLOF (46.8 AP) which were trained on significantly more data.
  • Speed: The model achieves 50 FPS, making it viable for real-time conveyor belt monitoring.
  • Compactness: With a model size of just 19MB, it is the smallest among all compared methods (for comparison, a standard Faster R-CNN is approximately 264MB).
  • Memory Efficiency: It utilizes significantly less inference memory (1059MB) compared to other FSOD methods like MPSR, which requires roughly 3821MB.

Discussion and Key Insights

The success of Faster OreFSDet stems from a focus on efficiency over complexity. The shift to CenterNet2 provides a significant speed advantage by reducing the computational burden of proposal generation. Furthermore, the authors identified that preserving spatial relationships is vital for distinguishing individual ore particles, which often overlap or adhere to one another. By aggregating features at two different scales via the DSA module, the detector avoids the common pitfall of missing small objects while tuned for larger ones.

Conclusion and Future Work

Faster OreFSDet demonstrates that it is possible to achieve high-accuracy, real-time object detection in data-scarce industrial environments. While currently focused on a single-class problem (ore), the framework provides a robust foundation for future developments.

The authors suggest that future research will involve extending the framework to Few-Shot Instance Segmentation to provide precise outlines of each ore particle, exploring One-Shot Detection capabilities, and developing new mechanisms to further strengthen the correlation between support and query features.

Summary updated

BibTeX
@article{zhang2023faster,
  author = {Yang Zhang and Le Cheng and Yuting Peng and Chengming Xu and Yanwei Fu and Bo Wu and Guodong Sun},
  title = {{Faster OreFSDet: A lightweight and effective few-shot object detector for ore images}},
  journal = {Pattern Recognition},
  year = {2023},
  url = {https://openreview.net/forum?id=cncFO1hdK8},
}
Figure