Low- and High-Fidelity Classifiers for Fast Road-Scene Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing systems for assisted and autonomous driving face challenges in achieving accurate and rapid object classification, detection, and location while managing computational intensity and maintaining accuracy.
Innovation Solution
A multi-stage image analysis system is implemented, utilizing a low-fidelity classifier and a high-fidelity classifier both based on Convolutional Neural Networks (CNNs). The low-fidelity classifier uses a down-sampled dataset to quickly identify candidate regions, while the high-fidelity classifier verifies these regions using high-resolution images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If Convolutional Neural Networks (CNNs) are used for image analysis, then classification accuracy is improved, but computational intensity increases
Solution Approach 1:
The patent applies segmentation by dividing the image analysis task into two distinct stages: a first CNN processes the entire image to identify candidate regions, while a second CNN processes only those specific candidate regions at higher resolution. This segmentation of the computational task reduces overall computational intensity while maintaining classification accuracy, as the computationally expensive second CNN is applied only to relevant portions of the image rather than the entire image.
2Measurement precision
If high-resolution images are processed for accurate classification, then measurement precision is improved, but processing time increases
Solution Approach 1:
The patent segments the processing workflow into two phases: rapid low-resolution screening of the entire image followed by focused high-resolution analysis of only candidate regions. This approach maintains measurement precision by applying high-resolution processing selectively, while dramatically reducing overall processing time compared to processing the entire high-resolution image.
Solution Approach 2:
The first CNN performs preliminary action by identifying candidate regions in the low-resolution image before the second CNN processes high-resolution versions of those regions. This preliminary screening step filters out non-candidate areas, ensuring that computationally intensive high-resolution processing is applied only where necessary, thus reducing total processing time while maintaining accuracy.
3Reliability
If comprehensive image analysis is performed to detect all objects, then reliability is improved, but productivity decreases
Solution Approach 1:
The patent improves productivity while maintaining detection reliability by segmenting the analysis into a fast first-pass CNN that screens the entire image, followed by a second CNN that verifies only candidate regions. This segmented approach processes images through multiple stages with increasing computational depth only where needed, thereby increasing overall processing throughput while maintaining comprehensive detection reliability.
Solution Approach 2:
The patent applies partial action by performing complete high-resolution analysis only on candidate regions identified by the first CNN, rather than applying exhaustive high-resolution analysis to the entire image. This partial processing approach maintains detection reliability for all objects while significantly improving processing throughput by avoiding unnecessary computation in non-candidate areas.
Data Source
AI summary
Disclosures herein teach applying a set of sections spanning a down-sampled version of an image of a road-scene to a low-fidelity classifier to determine a set of candidate sections for depicting one or more objects in a set of classes. The set of candidate sections of the down-sampled version may be mapped to a set of potential sectors in a high-fidelity version of the image. A high-fidelity classifier may be used to vet the set of potential sectors, determining the presence of one or more objects from the set of classes. The low-fidelity classifier may include a first Convolution Neural Network (CNN) trained on a first training set of down-sampled versions of cropped images of objects in the set of classes. Similarly, the high-fidelity classifier may include a second CNN trained on a second training set of high-fidelity versions of cropped images of objects in the set of classes.