Seeing Clearly and Detecting Precisely: Perceptual Enhancement and Focus Calibration for Small-Object Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Small object detection is difficult due to limited pixel data, unclear boundaries, and weak semantic information.
• Existing methods struggle with preserving fine details and precise localization, especially in cluttered or low-resolution environments.

Purpose of the Study:

• To introduce PEFC-Net, a novel framework designed to enhance both perceptual clarity and localization accuracy for small object detection.
• To address the loss of fine-grained features during downsampling and the lack of region-aware focus mechanisms in current approaches.

Main Methods:

• Developed the hybrid structural perception (HSP) module, utilizing wavelet decomposition and edge-aware refinement to encode spatial gradients and frequency components.
• Designed the axis-aligned focus calibration (AAFC) module, incorporating axis-sensitive pooling and shape-aware calibration for improved region-level focus and long-range contextual understanding.

Main Results:

• PEFC-Net demonstrated consistent outperformance over state-of-the-art methods across four challenging benchmarks: VisDrone-2019, TT100K, NWPU VHR-10, and DIOR.
• The framework showed robust performance in scenarios with occlusion, dense object distribution, and significant scale variation.

Conclusions:

• PEFC-Net effectively mitigates structural degradation and enhances region-aware focus for small object detection.
• The proposed method offers a significant advancement in accurately detecting small objects under challenging real-world conditions.