CSF-Net: Cross-Stage Fusion Network with Dual Backbone for Small Object Detection.

Area of Science:

• Computer Vision
• Object Detection

Background:

• Small object detection is challenging due to low pixel count and feature scarcity.
• Conventional single-backbone networks struggle to balance semantic extraction and detail preservation.
• Deep down-sampling and late-stage fusion hinder the recovery of lost edge and texture information.

Purpose of the Study:

• To propose an effective network architecture for small object detection.
• To address the limitations of conventional networks in capturing both semantic and fine-grained details.
• To enhance the performance of small target detection in computer vision applications.

Main Methods:

• Introduced the Cross-Stage Fusion Network (CSF-Net) with a dual backbone.
• Employed an asymmetric design with a shallow backbone for high resolution and a deep backbone for semantics.
• Utilized progressive cross-stage connections for early fusion of small object information.

Main Results:

• CSF-Net improved the mean Average Precision (mAP) of the YOLOV8 baseline from 62.8% to 67.0% on the Micro RGB UAV dataset.
• Demonstrated enhanced detection performance for small targets compared to the baseline.
• Validated the effectiveness of the proposed dual-backbone and cross-stage fusion approach.

Conclusions:

• The proposed CSF-Net effectively addresses the challenges of small object detection.
• The dual-backbone architecture and cross-stage fusion mechanism significantly improve detection accuracy.
• CSF-Net offers a promising solution for enhancing small object detection performance in computer vision.