Vehicle Neural Network Training With Image Variations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing methods for training neural networks to detect objects in vehicle environments are inefficient due to the high cost, time, and resource consumption associated with acquiring large datasets, which involves human judgment and is unreliable.
Innovation Solution
The method involves generating additional images through geometric transformations to increase the number of training images and corresponding ground truth data, allowing a memory augmented neural network to be trained with a smaller dataset, using convolutional and fully connected layers to process feature points, and selecting output results based on Mahalanobis distance and Kullback-Leibler divergence.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large datasets are acquired through human judgment for training neural networks, then object detection accuracy is improved, but training cost and time consumption increase significantly
Solution Approach 1:
The patent applies preliminary action by pre-processing training images through geometric transformations (rotation, flipping, scaling) to generate multiple variations of each image. This preprocessing step creates an expanded training dataset before the neural network training begins, allowing the model to learn from more diverse examples without requiring additional manual annotation time. The transformed images are stored and reused during training, significantly reducing the time needed to acquire and process large datasets while maintaining high object detection accuracy.
2Measurement precision
If large datasets are acquired through human judgment for training neural networks, then object detection accuracy is improved, but resource consumption increases
Solution Approach 1:
The patent applies copying by creating multiple copies of each training image through geometric transformations. Instead of requiring a large number of unique, manually annotated images, the system generates multiple transformed versions (rotated, flipped, scaled) of a smaller set of original images. These copied and transformed images serve as additional training examples, reducing the need to process and store large volumes of original data while still providing sufficient diversity for accurate object detection.
3Productivity
If geometric transformations are applied to generate additional training images, then training efficiency is improved, but computational complexity increases
Solution Approach 1:
The patent applies segmentation by dividing the image transformation process into discrete, manageable operations. Each training image undergoes specific geometric transformations (rotation by fixed angles, horizontal/vertical flipping, scaling by fixed factors) rather than continuous transformations. This segmented approach to image processing reduces computational complexity by using a有限的 set of predetermined transformation parameters, making the preprocessing step more efficient while still generating sufficient image variations for effective neural network training.
Data Source
AI summary
A computer, including a processor and a memory, the memory including instructions to be executed by the processor to train a neural network included in a memory augmented neural network based on one or more images and corresponding ground truth in a training dataset by transforming the one or more images to generate a plurality of one-hundred or more variations of the one or more images including variations in the ground truth and process the variations of the one or more images and store feature points corresponding to each variation of the one or more images in memory associated with the memory augmented neural network. The instructions can include further instructions to process an image acquired by a vehicle sensor with the memory augmented neural network, including comparing a feature variance set for the image acquired by the vehicle sensor to the stored processing parameters for each variation of the one or more images, to obtain an output result.