CNN Biometric Detection with Orientation and Objectivity Scoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional biometric systems using convolutional neural networks (CNNs) face difficulties in detecting multiple biometric traits in complex images, such as a hand with multiple fingers, due to incorrect detection of traces, phalanges, and orientation issues, which are not robustly handled by existing algorithms.
Innovation Solution
A method employing a CNN that generates representation vectors for candidate regions of interest in an input image, including position, size, orientation, and objectivity scores, with a filtering process to select and correct the orientation of biometric traits, using a cost function that includes regression terms for position, size, objectivity, and orientation, to improve detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional CNN detection algorithms are used to identify biometric traits in complex images, then the detection process is simple and fast, but the accuracy deteriorates due to incorrect detection of traces, phalanges, and orientation issues
Solution Approach 1:
The algorithm segments the detection process into multiple stages: initial candidate region proposal, filtering based on objectivity scores, and final selection of biometric traits. This segmentation allows the system to handle complex images by breaking down the detection task into manageable steps, improving accuracy without overwhelming computational complexity
Solution Approach 2:
The algorithm performs preliminary filtering of candidate regions using objectivity scores before final biometric trait identification. By pre-processing and filtering potential regions of interest before detailed analysis, the system reduces false detections of traces and phalanges, improving detection accuracy while maintaining efficiency
2Reliability
If existing CNN algorithms are used for biometric detection, then the system remains simple, but reliability deteriorates due to non-robust handling of orientation variations and multiple traits
Solution Approach 1:
The algorithm incorporates orientation parameters and objectivity scores as additional dimensions for evaluating candidate regions. By changing the parameter space to include orientation robustness and confidence metrics, the system reliably handles varied orientations and distinguishes true biometric traits from false positives like traces and phalanges
Solution Approach 2:
The algorithm adds a new dimension of evaluation by computing objectivity scores for each candidate region. This additional dimensional assessment layer enables the system to filter unreliable detections and maintain robustness across different orientations and image complexities without requiring fundamentally new algorithmic approaches
3Measurement precision
If detailed filtering and orientation correction are implemented, then detection accuracy improves, but processing time increases
Solution Approach 1:
The algorithm applies partial filtering by computing objectivity scores for all candidate regions but only performing detailed orientation correction and selection on high-scoring candidates. This partial application of complex processing steps maintains detection accuracy while avoiding unnecessary computation on low-probability regions, reducing overall processing time
Data Source
AI summary
A method for detecting at least one biometric trait visible in an input image, by means of a convolutional neural network, the method wherein it comprises the implementation, by data processing means of a client, of steps of: (a) Generating, by means of a feature extraction block of said CNN, a plurality of representation vectors each defining a candidate region of interest of said input image potentially containing a biometric trait, the representation vector of a candidate region of interest comprising at least one position value of the candidate region of interest, at least one size value of the candidate region of interest, an orientation value of the candidate region of interest, and an objectivity score of the candidate region of interest; (b) Selecting, by means of a filtering block of said CNN, at least one region of interest from said candidate regions based on the representation vectors thereof.