Frequency Domain Object Detection via Fourier Transform
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Object detection methods in images, such as face detection and pedestrian detection, are computationally expensive, leading to high processing latency, which is impractical for mobile devices and other computer systems.
Innovation Solution
The approach involves converting image and object detection filters into frequency domain representations, allowing for the generation of a composite frequency domain representation through convolution or point-wise multiplication, which significantly reduces the computational cost and processing time by leveraging the equivalence of convolution in the spatial domain to multiplication in the frequency domain.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If sliding-window-based object detection method is used, then object detection accuracy is maintained, but processing time and computational cost increase significantly
Solution Approach 1:
The patent transforms the object detection problem from spatial domain to frequency domain using Fourier transforms. By converting images and filters into frequency representations, the convolution operation becomes simple element-wise multiplication, dramatically reducing computational complexity from O(kdxymn) to O(kdxy) while preserving detection accuracy
Solution Approach 2:
The patent replaces the traditional sliding-window convolution mechanism with frequency domain multiplication. This substitution leverages the mathematical equivalence between spatial convolution and frequency multiplication, eliminating the need for exhaustive sliding-window operations while maintaining the same detection functionality
2Reliability
If sliding-window-based object detection method is used, then comprehensive object detection is achieved, but computational complexity increases to O(kdxymn)
Solution Approach 1:
The patent applies Fourier transformation to convert the detection problem into frequency domain, where the complex sliding-window convolution operations are replaced by simple element-wise multiplication of frequency representations, reducing computational complexity from O(kdxymn) to O(kdxy)
Solution Approach 2:
The patent changes the domain parameter from spatial to frequency, transforming the image and filter representations. This parameter change fundamentally alters the computational operation from convolution to multiplication, achieving the same detection reliability with significantly reduced complexity
Data Source
AI summary
In an embodiment, a method comprises obtaining a frequency domain representation associated with an image; obtaining one or more frequency domain representations of one or more object detection filters; generating a composite frequency domain representation based on the frequency domain representation associated with the image and the one or more frequency domain representations of the one or more object detection filters; and detecting one or more objects in the image based on the composite frequency domain representation. The frequency domain representation associated with the image may be obtained based on a forward transform performed on an image feature description. The image feature description may be obtained based on a feature extraction performed on the image. The one or more frequency domain representations of the one or more object detection filters may be obtained based on one or more Fourier transforms performed on the one or more object detection filters.