Vehicle Training Image Selection Using Pixel and Class Entropy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning models for autonomous driving are not accurately predicting vehicle paths due to insufficient and condition-specific training image collection, leading to potential safety risks and inefficiencies in improving model performance.
Innovation Solution
An apparatus and method for collecting training images using a vehicle-mounted camera sensor to determine pixel and class entropy, prioritizing images based on entropy ranges, and periodically transmitting them to an external server to enhance model performance without disrupting vehicle operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If training images are collected under specific conditions set by a user, then the collection process is simple and controllable, but the model performance cannot be improved and significant time is spent on setting conditions
Solution Approach 1:
The system automatically determines entropy of each pixel and class in captured images without user intervention, and autonomously selects training images based on preset entropy ranges. This self-service mechanism eliminates the need for users to manually set collection conditions while ensuring optimal model performance through objective entropy-based selection criteria.
Solution Approach 2:
The patent introduces entropy as a new parameter for evaluating and selecting training images. By calculating entropy of pixels and classes in captured images, the system objectively determines image quality and diversity, replacing subjective user-defined conditions with a quantifiable parameter that directly correlates with model training effectiveness.
2Ease of manufacture
If training images are collected only under user-set conditions, then the collection process is straightforward, but training images cannot be collected under undefined conditions
Solution Approach 1:
The entropy-based selection criterion serves multiple functions simultaneously: it evaluates image quality, ensures class diversity, and adapts to various driving scenarios without requiring condition-specific configurations. This universal approach allows the system to collect training images across all possible conditions while maintaining consistent quality standards.
Solution Approach 2:
The system autonomously determines whether to collect an image based on entropy calculations, eliminating the need for predefined collection conditions. This self-service mechanism enables the system to adaptively collect training images under any condition while maintaining quality through objective entropy-based filtering.
3Speed
If existing technology predicts vehicle path using only shape information and driving information, then the prediction process is simple and fast, but the prediction accuracy is insufficient
Solution Approach 1:
Entropy serves as an intermediary parameter that bridges the gap between simple image capture and accurate path prediction. By using entropy to select training images with high information content and class diversity, the system prepares better training data that enables more accurate path prediction models while maintaining efficient processing speeds.
Solution Approach 2:
The system performs preliminary entropy-based filtering and selection of training images before model training. This preliminary action ensures that only high-quality, diverse images are used for training, which significantly improves prediction accuracy without adding computational overhead during the actual path prediction process.
Data Source
AI summary
Provided are an apparatus for collecting a training image of a deep learning model and a method for the same. The apparatus may include a camera sensor configured to capture an image of a surrounding of a vehicle and a controller configured to determine an entropy of each pixel in the image. The controller may also be configured to determine an entropy of each class in the image based on the entropy of each pixel in the image and determine whether to collect the image based on the entropy of each class.