Neural Network Training for Low-Confidence Out-of-Distribution Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing vehicle recognition systems in average speed monitoring face challenges in accurately distinguishing between vehicle number plates and other signs, leading to incorrect vehicle matching and speed calculations, particularly when signs like warning signs are mistaken for number plates.
Innovation Solution
A method for training an artificial neural network (ANN) by incorporating a first set of inputs within the desirable input distribution (vehicle number plates) and a second set of inputs outside this distribution (predetermined vehicle signs) with randomly assigned labels, reducing the network's confidence in misclassifying non-number plate signs by defining additional decision hyperplanes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If an optical character recognition algorithm is used to recognise vehicles, then the system can automatically match vehicles between locations, but the system may mistakenly match vehicle signs (e.g., warning signs) as number plates, leading to false vehicle pairing
Solution Approach 1:
The system performs preliminary classification of image regions to identify whether they contain number plates or other signs before attempting character recognition. By pre-sorting inputs and applying different processing paths, the system avoids misclassifying warning signs as number plates while maintaining efficient automatic recognition for genuine number plates
Solution Approach 2:
The image processing is segmented into distinct stages: first identifying potential number plate regions, then classifying them as either actual number plates or other signs, and finally applying character recognition only to confirmed number plates. This segmentation allows the system to maintain high productivity for valid number plates while eliminating false matches from vehicle signs
2Speed
If the system increases confidence in classifying all detected characters, then faster vehicle matching is achieved, but misclassification of non-number plate signs increases
Solution Approach 1:
The system applies different confidence thresholds and classification criteria depending on the local context of the detected region. For regions identified as genuine number plates, the system uses standard recognition confidence thresholds. For regions identified as other signs (warning signs, safety signs), the system applies different evaluation criteria that prevent high-confidence misclassification, thus maintaining speed for valid matches while improving precision for distinguishing signs from number plates
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
A method for an artificial neural network is provided. The method comprises providing the artificial neural network, wherein the artificial neural network is trained to reduce the confidence of the artificial neural network in classifying inputs outside a desirable input distribution is provided by training the artificial neural network (230) using a first set of inputs and associated labels, the labels correctly classifying the corresponding inputs, wherein the inputs within the first set are within the desirable input distribution of the artificial neural network; training the artificial neural network using a second set of inputs and associated labels, wherein the inputs within the second set are outside of the desirable input distribution of the artificial neural network, and wherein the labels within the second set comprise randomly assigned classifications; and feeding one or more inputs through the artificial neural network to determine classifications, classifying the inputs, and confidence values associated with the classifications.