Object Recognition via Block Average Value Assignment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object recognition technologies face challenges with high computational overhead, convergence issues during learning, pose classification failures, and increased calculation complexity as the number of training sets grows, particularly in handling variations and multi-class learning processes.
Innovation Solution
The method involves setting calculation blocks in images to acquire average or adjusted values, matching these with reference blocks of varying positions and sizes, and using the assigned values to recognize objects by comparing them across windows, thereby reducing computational load and improving classification performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If binary descriptor methods (LBP, MCT) are used for object recognition, then calculation speed is improved, but classification performance deteriorates due to convergence failures and pose classification issues
Solution Approach 1:
The patent changes the parameter representation from direct pixel intensity comparisons (LBP, MCT) to average intensity values of block regions. This parameter transformation allows the method to maintain fast binary code generation while improving convergence during learning processes and achieving better pose classification performance by using more robust average values that are less sensitive to local variations
2Reliability
If the number of training sets is increased to improve recognition accuracy, then classification performance is improved, but calculation complexity increases sharply
Solution Approach 1:
The patent divides the image into multiple blocks and calculates average values for each block separately. This segmentation approach allows efficient processing of large training sets by breaking down complex global comparisons into simpler local block comparisons, thereby maintaining low calculation complexity even when the number of training sets increases
3Measurement precision
If pixel intensity comparisons are performed between distant pixels (Ferns method) to improve feature extraction, then feature discrimination is improved, but recognition rate drops significantly when pixel values contain errors
Solution Approach 1:
The patent changes from comparing individual pixel intensities (Ferns method) to comparing average intensity values of blocks. This parameter change makes the feature extraction more robust to pixel-level errors and noise, as the average value smooths out local variations and errors, thereby maintaining high recognition rates even when individual pixel values are inaccurate
Data Source
AI summary
A method for supporting object recognition is disclosed. The method includes the steps of: setting calculation blocks, each of which includes one or more pixels in an image, acquiring respective average values of the pixels included in the respective calculation blocks, and matching information on the respective calculation blocks with the respective average values or respective adjusted values derived from the respective average values; referring to information on windows, each of which includes information on one or more reference blocks which are different in at least either positions or sizes and information on corresponding relations between the calculation blocks and the average values or the adjusted values, to thereby assign the respective average values or the respective adjusted values to the respective reference blocks; and acquiring necessary information by using the respective average values or the respective adjusted values assigned to the respective reference blocks.