MBAN: multi-branch attention network for small object detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Small object detection is challenging due to low resolution and information loss during downsampling in complex scenes.
• Existing methods struggle with accurately detecting and localizing small objects.

Purpose of the Study:

• To propose a novel Multi-branch Attention Network (MBAN) to enhance small object detection performance.
• To address the information loss and feature representation issues specific to small objects.

Main Methods:

• Developed a Multi-branch Attention Module (MBAM) combining a multi-branch structure (convolution, maxpooling) and the SimAM attention mechanism.
• Introduced Adaptive Clustering Relocation (ACR) as a pre-processing method for improved small object localization.
• Integrated MBAM and ACR into the proposed Multi-branch Attention Network (MBAN).

Main Results:

• MBAN demonstrated significant performance improvements over existing algorithms on benchmark datasets.
• Achieved a mean Average Precision (mAP) of 96.55% on the NWPU VHR-10 dataset.
• Achieved a mean Average Precision (mAP) of 84.96% on the PASCAL VOC dataset.

Conclusions:

• The proposed MBAN effectively enhances small object detection by reducing parameter count and information loss.
• MBAN significantly improves the representation of small object features, leading to superior detection accuracy.
• The study validates MBAN's effectiveness on standard datasets, proving its capability in challenging small object detection scenarios.