NPU Model Switching for Object Detection Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing electronic devices equipped with artificial intelligence semiconductors face challenges in maintaining accurate object detection and tracking, particularly in varying environments and conditions such as different altitudes and object sizes.
Innovation Solution
A neural processing unit (NPU) is mounted on movable devices, featuring multiple processing elements (PEs) that can selectively perform operations of different artificial neural network models based on determination data, including object detection performance data and environmental conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single artificial neural network model is used for object detection, then the device complexity is low, but the object detection accuracy varies significantly across different environmental conditions and object sizes
Solution Approach 1:
The system dynamically switches between different artificial neural network models based on environmental conditions and object characteristics. The controller selects which model to execute by evaluating determination data including object size, detection altitude, and environmental factors, allowing the object detection accuracy to be optimized for varying conditions without permanently increasing device complexity
Solution Approach 2:
The system changes operational parameters by switching between multiple pre-configured neural network models, each optimized for specific parameter ranges such as object size, altitude, and environmental conditions. This allows the detection system to adapt to different scenarios by selecting the appropriate model parameters rather than using a single fixed model
2Measurement precision
If multiple artificial neural network models are maintained ready for immediate execution, then the object detection accuracy across various conditions is improved, but the power consumption increases
Solution Approach 1:
Instead of maintaining multiple models in active ready states, the system dynamically loads and executes only the specific neural network model needed for the current environmental conditions and object characteristics. This on-demand model execution significantly reduces power consumption compared to keeping multiple models continuously active, while still providing accurate detection across varying conditions
Solution Approach 2:
The system extracts and executes only the specific neural network model portion needed for the current detection task based on determination data. Rather than maintaining all models in memory and active state, only the relevant model is loaded and executed, reducing the overall power consumption while maintaining detection accuracy
3Adaptability or versatility
If the neural network model is switched frequently based on environmental conditions, then the adaptability to various detection scenarios is improved, but the processing time and decision-making delay increase
Solution Approach 1:
Multiple neural network models are pre-configured and prepared in advance, each optimized for specific environmental conditions and object characteristics. The determination data is pre-evaluated to identify which pre-configured model should be executed, eliminating the need for complex real-time model generation or selection algorithms, thus reducing processing time while maintaining high adaptability
Solution Approach 2:
The system uses predetermined parameter thresholds and decision criteria to quickly determine which pre-configured model to execute based on determination data such as object size, altitude, and environmental factors. This parameter-based decision-making approach enables rapid model selection with minimal processing time, balancing adaptability with speed
Data Source
AI summary
A neural processing unit (NPU) mounted on a movable device for detecting object is provided. The NPU may comprise a plurality of processing elements (PEs), configured to process an operation of a first artificial neural network model (ANN) and an operation of a second ANN different from the first ANN; a memory configured to store a portion of a data of the first ANN and the second ANN; and a controller configured to control the PEs and the memory to selectively perform a convolution operation of the first ANN or the second ANN based on a determination data, wherein the determination data may include an object detection performance data of the first ANN and the second ANN, respectively.