An Enhanced Region Proposal Network for object detection using deep learning method.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Faster Region-based Convolutional Network (Faster R-CNN) is a leading object detection technique.
• The Region Proposal Network (RPN) component of Faster R-CNN has limitations in object detection accuracy.

Purpose of the Study:

• To introduce an Enhanced Region Proposal Network (ERPN) to improve object detection performance.
• To address the shortcomings of existing RPNs in generating high-quality region proposals.

Main Methods:

• Developed a deconvolutional feature pyramid network (DFPN) for enhanced region proposal quality.
• Designed novel anchor boxes with interspersed scales and adaptive aspect ratios for better localization.
• Implemented a particle swarm optimization-based support vector machine (PSO-SVM) for improved anchor box classification.
• Enhanced the classification loss component of the RPN's multi-task loss function.

Main Results:

• ERPN achieved state-of-the-art performance, with VGG-16 obtaining 78.6% mAP on PASCAL VOC 2007, 74.4% on VOC 2012, and 31.7% on COCO.
• The method demonstrated superior detection speed at 5.8 frames per second (fps).
• ERPN showed significant effectiveness in detecting small objects.

Conclusions:

• The proposed ERPN significantly enhances object detection capabilities compared to existing methods.
• ERPN's improvements in proposal generation, localization, and classification contribute to its superior performance.
• The method is particularly effective for challenging scenarios like small object detection.