ssFPN: Scale Sequence (S²) Feature-Based Feature Pyramid Network for Object Detection.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Object detection is crucial in computer vision, with Convolutional Neural Networks (CNNs) and Feature Pyramid Networks (FPNs) improving accuracy.
• Existing FPNs struggle with detecting small objects due to information loss in deeper CNN layers.
• Small object detection remains a challenge, impacting overall detection performance.

Purpose of the Study:

• To propose a new FPN model, the scale sequence (S2) feature-based FPN (ssFPN), for improved multi-scale object detection.
• To introduce and extract a novel scale sequence (S2) feature using 3D convolution on FPN levels.
• To enhance the detection of small objects by strengthening their information content.

Main Methods:

• Proposed a scale sequence (S2) feature extracted via 3D convolution on FPN levels, inspired by scale-space theory.
• Developed a feature-level super-resolution approach to demonstrate the S2 feature's efficiency.
• Integrated the S2 feature into various object detection models (Faster R-CNN, Mask R-CNN, YOLO series) for evaluation.

Main Results:

• The S2 feature improved AP by up to 1.6% for Faster R-CNN and 1.4% for Mask R-CNN on the MS COCO dataset.
• Small object detection accuracy (APS) saw improvements of up to 1.2% and 1.1% for these models, respectively.
• YOLO series models also showed AP gains, with notable improvements in small object detection.

Conclusions:

• The proposed S2 feature effectively enhances object detection, especially for small and multi-scale objects.
• The S2 feature is versatile and can be integrated into existing FPN-based object detection frameworks.
• The feature-level super-resolution experiments confirmed the S2 feature's capability to boost classification accuracy on low-resolution images.