Real-Time Object Detection With Reduced Region Proposal Network via Multi-Feature Concatenation.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Object detection is crucial in deep learning, with one-stage and two-stage neural networks being primary architectures.
• Faster R-CNN, a popular two-stage detector, faces challenges in achieving real-time inference without accuracy loss.

Purpose of the Study:

• To improve the inference time of the Faster R-CNN architecture for real-time object detection.
• To maintain or enhance detection accuracy while significantly reducing computational load.

Main Methods:

• Applied pruning to reduce weights in convolutional and fully connected (FC) layers for computation efficiency.
• Introduced a reduced region proposal network (RRPN) incorporating dilated convolution and multi-scale feature concatenation.
• Developed intra-layer concatenation and proposal refinement for efficient feature map integration.

Main Results:

• Achieved significant parameter compression: 81.2% for ZF-Net and 73% for VGG16.
• Reduced computation by 66% (ZF-Net) and 77% (VGG16).
• Increased inference speed to 40 FPS (ZF-Net) and 27 FPS (VGG16).
• Enhanced accuracy to 60.2% mAP (ZF-Net) and 69.1% mAP (VGG16).

Conclusions:

• The proposed RRPN with feature integration effectively compensates for accuracy loss from pruning.
• The method enables real-time object detection with improved speed and accuracy.
• Demonstrated successful application on PASCAL VOC datasets using ZF-Net and VGG16 backbones.