ANN Feature Map Encoding for Vehicle Image Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for processing image information in neural networks, such as convolutional neural networks (CNNs), face challenges with high computational and storage requirements, especially in resource-constrained environments like vehicle-mounted computing units. This is exacerbated by the need for one-hot encoding in classification tasks and the linear increase in feature maps with the number of classes or quantization steps.
Innovation Solution
The proposed method employs an artificial neural network with an encoder and a decoder, configured to output multiple feature maps that encode image region allocations to classes or quantized numerical output information. This encoding reduces the number of feature maps below the number of classes or quantization steps, achieving information compression and lowering computational and storage demands.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If one-hot encoding is used for output representations in classification tasks, then class information can be clearly represented, but the number of feature maps increases linearly with the number of classes, leading to high computational and storage requirements
Solution Approach 1:
The patent changes the encoding parameter from one-hot encoding (requiring N feature maps for N classes) to binary encoding (requiring only log2(N) feature maps). This parameter change in the encoding scheme reduces the number of feature maps from linear to logarithmic growth with the number of classes, directly resolving the contradiction between accurate class representation and computational complexity.
2Measurement precision
If the number of feature maps is increased to represent more classes or quantization steps, then classification or regression accuracy is improved, but computational cost and storage capacity requirements increase proportionally
Solution Approach 1:
The patent applies binary encoding to change the parameter representation from decimal (one-hot) to binary format. This allows the neural network to represent N classes or quantization steps using only log2(N) feature maps instead of N feature maps, reducing computational cost while maintaining the ability to distinguish all N categories with sufficient accuracy.
3Measurement precision
If pixel-based methods are used for semantic segmentation with high number of classes, then detailed image region classification is achieved, but high storage capacities are required which is problematic for vehicle-mounted computing units
Solution Approach 1:
The patent changes the output encoding parameter from one-hot encoding to binary encoding in pixel-based semantic segmentation. This reduces the storage requirement per pixel from N bytes (for N classes) to log2(N) bytes, enabling detailed image region classification while fitting within the storage constraints of vehicle-mounted computing units.
Data Source
AI summary
A method for processing image information of an imaging sensor of a vehicle in an artificial neural network (“ANN”) is disclosed. The ANN includes at least one encoder and one decoder. The ANN solves a classification task with a plurality of classes and/or a regression task, in which numerical output information quantized according to a plurality of quantization steps is provided. The ANN outputs multiple feature maps at the output interface, wherein allocations of image regions of the image information to classes or numerical output information quantized regarding the image information is/are output by the feature maps in an encoded manner.