Object Detection Using DCT Frequency Domain Whitening
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current face detection technologies face challenges such as poor accuracy due to lack of discriminating facial features, excessive data leading to slow processing, inadequate consideration of local pixel structures, uneven illumination, and varying viewpoints.
Innovation Solution
The method involves selecting neighborhoods around each pixel, weighting pixel values using whitened matrices obtained through techniques like local binary pattern, local phase quantization, or modified census transform, and verifying object features across multiple pixels to enhance detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If simple Haar-wavelet features are used for face detection, then processing speed is improved, but detection accuracy deteriorates due to lack of discriminating facial features
Solution Approach 1:
The patent transforms pixel data from spatial domain to frequency domain using Discrete Cosine Transform (DCT), changing the representation parameters of facial features. This allows extraction of discriminative frequency-domain features that maintain processing efficiency while improving detection accuracy beyond simple spatial features like Haar-wavelets
Solution Approach 2:
The patent replaces traditional mechanical/image-processing approaches (Haar-wavelets, LBP) with a frequency-domain transformation approach using DCT. This substitution enables more effective feature extraction by operating in a different mathematical domain, achieving both speed and accuracy
2Measurement precision
If extensive feature extraction is performed to improve accuracy, then detection precision is improved, but processing time increases due to excessive data
Solution Approach 1:
The patent extracts only the most discriminative frequency coefficients from the DCT-transformed image data, rather than processing all pixel information. By selecting and extracting only the essential frequency components that contain facial feature information, the method achieves high accuracy with reduced processing time
Solution Approach 2:
The patent segments the frequency domain into different coefficient groups, processing and analyzing them separately. This segmentation allows the system to focus computational resources on the most informative frequency bands, improving efficiency while maintaining accuracy
3Device complexity
If traditional feature techniques are used, then device complexity is reduced, but reliability deteriorates due to poor consideration of local pixel structure and varying illumination
Solution Approach 1:
The patent changes from spatial-domain parameters to frequency-domain parameters using DCT transformation. This parameter transformation inherently handles illumination variations and enhances local structural information, improving reliability without significantly increasing system complexity
Solution Approach 2:
The patent applies DCT transformation as a preliminary step before feature extraction and classification. This preliminary frequency-domain transformation pre-processes the data to make it more robust to illumination changes and viewpoint variations, improving reliability for subsequent detection steps
Data Source
AI summary
There is provided an improved solution for detecting and tracking objects in digital images. The solution comprises selecting a neighborhood for each pixel under observation, the neighborhood being of known size and form, and reading pixel values of the neighborhood. Further the solution comprises selecting at least one set of coefficients for weighting each neighborhood such that each pixel value of each neighborhood is weighted with at least one coefficient; searching for an existence of at least one object feature at each pixel under observation on the basis of a combination of weighted pixel values at each neighborhood; and verifying the existence of the object in the digital image on the basis of the searches of existence of at least one object feature at a predetermined number of pixels.