Enhancing Object Detection With Fourier Series

Area of Science:

• Computer Vision
• Image Analysis
• Machine Learning

Background:

• Traditional object detection models struggle to retain intricate object outline details.
• Loss of fine contour information limits performance, especially for non-standard object shapes.

Purpose of the Study:

• To introduce a novel object detection method that preserves detailed object outline information.
• To improve feature extraction and descriptive capabilities for complex object geometries.

Main Methods:

• Developed Fourier Series Object Detection (FSD) which encodes object outlines using one-dimensional periodic Fourier series.
• Constructed a Fourier Series Model (FSM) to regress Fourier series for objects.
• Implemented Rolling Optimization Matching for Fourier loss to stabilize training.

Main Results:

• Achieved AP50 of 73.3% on the DOTA 1.5 dataset, exceeding the state-of-the-art by 6.44%.
• Attained AP50 of 97.25% on the UCAS dataset, surpassing existing methods.
• Demonstrated improved feature extraction for non-rectangular and elongated objects.

Conclusions:

• FSD effectively retrieves detailed object outlines, enhancing detection accuracy.
• Introduced Fourier power spectrum and Fourier vector for richer semantic scene representation.
• Paved a new direction for the evolution of object detection methodologies.