Object Detection Based on Faster R-CNN Algorithm with Skip Pooling and Fusion of Contextual Information.

Area of Science:

• Computer Vision
• Deep Learning
• Object Detection

Background:

• Faster region-based convolutional neural network (Faster R-CNN) is a leading object detection method.
• Standard Faster R-CNN struggles with object occlusion, deformation, and small sizes.

Purpose of the Study:

• To enhance Faster R-CNN's object detection performance in challenging scenarios.
• To improve accuracy and efficiency for occluded, deformed, or small objects.

Main Methods:

• Incorporated a context information feature extraction model.
• Implemented skip pooling to capture richer contextual object information.
• Replaced the region proposal network (RPN) with a guided anchor RPN (GA-RPN) for improved feature utilization.

Main Results:

• Achieved an average improvement of 6.857% in mean average precision (mAP) compared to baseline Faster R-CNN and other algorithms like YOLOv3 and SSD512.
• Maintained a competitive recall rate while boosting detection performance.
• Demonstrated superior detection rates and efficiency in special conditions.

Conclusions:

• The proposed enhanced Faster R-CNN algorithm effectively addresses limitations in object detection.
• The integration of contextual information, skip pooling, and GA-RPN significantly boosts performance for difficult object detection cases.
• This method offers a more robust and efficient solution for real-world object detection challenges.