Grid-Based CNN Parameter Switching for Atypical Road Layouts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional Neural Networks (CNN) used in autonomous driving struggle when input images deviate from the typical arrangement used in training, leading to inefficient processing due to optimized parameters being specific to center-located roads.
Innovation Solution
A method utilizing a grid generator to divide images into sections by using a grid generator to determine parameters for testing, which includes steps of acquiring a test image, determining object and non-object locations, generating subsections, and applying neural network operations using parameters learned from similar class information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If CNN parameters are optimized for images with roads in centers, then processing efficiency is improved for typical arrangements, but processing accuracy deteriorates when image arrangements differ from training
Solution Approach 1:
The patent divides the input image into multiple regions (e.g., center region, left region, right region) and applies different pre-trained CNN parameters to each region based on its characteristics. This segmentation allows the system to maintain high processing efficiency through specialized parameters while achieving accurate detection across diverse image arrangements by selecting appropriate parameters for each region.
Solution Approach 2:
The patent dynamically selects and switches between different sets of CNN parameters based on the detected arrangement characteristics of the input image. The system determines whether the image follows a typical arrangement (road in center) or atypical arrangement (road on side), and accordingly switches between corresponding parameter sets, enabling adaptive processing that maintains both efficiency and accuracy.
2Speed
If CNN uses fixed parameters from training, then processing speed is maintained, but adaptability to different image arrangements deteriorates
Solution Approach 1:
The patent pre-trains multiple sets of CNN parameters during the training phase, each optimized for specific image arrangements (e.g., typical arrangement with road in center, atypical arrangements with road on sides). During inference, the system directly selects from these pre-prepared parameter sets based on the input image characteristics, avoiding time-consuming retraining while maintaining high adaptability to different arrangements.
Solution Approach 2:
The patent changes the parameters of the neural network based on the detected image arrangement characteristics. The system selects different parameter sets from pre-trained models according to whether the image exhibits typical or atypical arrangements, enabling the same CNN architecture to adapt to various scenarios without modifying its structural configuration.
3Ease of manufacture
If CNN is trained on typical arrangements only, then training simplicity is improved, but generalization to atypical arrangements deteriorates
Solution Approach 1:
The patent creates a universal CNN system that can handle multiple image arrangements by training separate parameter sets for different scenarios (typical and atypical arrangements). Each parameter set is specialized for its intended arrangement type, yet the overall system achieves multi-functionality by selecting the appropriate parameter set based on input characteristics, eliminating the need for complex retraining or single-parameter-set compromises.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method of neural network operations by using a grid generator is provided for converting modes according to classes of areas to satisfy level 4 of autonomous vehicles. The method includes steps of: a computing device (a) instructing a detector to acquire object location information for testing and class information; (b) instructing the grid generator to generate section information by referring to the object location information for testing; (c) instructing a neural network to determine parameters for testing, to be used for applying the neural network operations to either (i) the subsections including each of the objects for testing and each of non-objects for testing, or (ii) each of sub-regions, in each of the subsections, where said each of the non-objects for testing is located; and (d) instructing the neural network to apply the neural network operations to the test image for testing to thereby generate neural network outputs.