Generic Feature Extraction Unit for Multi-Task Neural Network Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current recognition systems that use neural networks for vehicle control tasks, such as sign identification, pedestrian detection, and white-line detection, require a large amount of learning data and multiple neural networks, leading to high development and updating costs due to similar feature extraction being performed for multiple tasks with independent networks.
Innovation Solution
A recognition system with a generic-feature extraction unit using a generic neural network that extracts common features for multiple recognition tasks, reducing the need for multiple neural networks and allowing for shared learning data sets, and enabling low-power, high-speed feature extraction with hardware implementation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If multiple independent neural networks are prepared for different recognition tasks, then each task can be performed with dedicated processing, but development costs and computing resources increase significantly
Solution Approach 1:
The patent implements a universal neural network that can perform multiple recognition tasks (sign identification, pedestrian detection, white-line detection) by dynamically switching between different recognition models stored in memory. This single multi-functional network replaces the need for multiple dedicated neural networks, reducing development costs and computing resources while maintaining task-specific performance through model selection.
2Reliability
If similar feature extraction is performed for multiple recognition tasks with independent neural networks, then each task receives dedicated processing, but development and updating costs increase
Solution Approach 1:
The patent merges the feature extraction functionality into a single shared component that is common across all recognition tasks. Instead of having separate feature extraction networks for each task, the system uses one unified feature extraction mechanism that feeds into different recognition models, thereby reducing development and updating costs while maintaining task-specific accuracy through separate recognition model processing.
3Device complexity
If a single generic neural network is used for multiple recognition tasks, then development costs are reduced, but processing speed and power consumption may decrease
Solution Approach 1:
The patent implements a dynamic system where the neural network can switch between different recognition models based on the current task requirements. This dynamic model selection allows the system to optimize processing speed and power consumption by loading only the necessary models into memory and using them when needed, rather than maintaining all models simultaneously, thus achieving both cost reduction and performance optimization.
Data Source
AI summary
A recognition system includes: a sensor processing unit (SPU) that performs sensing to output a sensor value; a task-specific unit (TSU) including an object detection part that performs an object detection task based on the sensor value and a semantic segmentation part that performs a semantic segmentation task based on the sensor value; and a generic-feature extraction part (GEU) including a generic neural network disposed between the sensor processing unit and the task-specific unit, the generic neural network being configured to receive the sensor value as an input to extract a generic feature to be input in common into the object detection part and the semantic segmentation part.