Parallel Target Detection Models for Low-Load Image Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning models for target detection face challenges in being applied to devices with limited computation resources, leading to either inadequate performance or increased processing demands when layer or bit count is reduced.
Innovation Solution
Employing multiple pre-trained models specialized for different target sizes and characteristics, enabling parallel processing to enhance detection accuracy while reducing overall processing load.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the number of layers of analysis model or number of bits for computation is increased, then detection accuracy is improved, but processing amount increases
Solution Approach 1:
The patent divides the detection task into multiple parallel processing paths, each handling different target sizes. Instead of using a single complex model with many layers, the system segments the problem into multiple simpler models that process simultaneously, reducing the processing burden on each individual model while maintaining overall detection accuracy.
Solution Approach 2:
The patent applies partial action by using multiple models with different complexity levels tailored to specific detection needs. Each model is optimized for particular target size ranges, allowing the system to use appropriate computational resources for each detection task rather than applying excessive computation uniformly across all detection scenarios.
2Device complexity
If the number of layers of analysis model or number of bits for computation is reduced, then processing amount is decreased, but detection accuracy deteriorates
Solution Approach 1:
The patent merges multiple simple models into a unified parallel processing system. By combining several models with fewer layers into a coordinated ensemble that processes images simultaneously, the system achieves detection accuracy comparable to or exceeding that of single complex models while reducing the processing requirements of each individual model.
Solution Approach 2:
The patent changes parameters by training multiple models with different architectures optimized for specific target size ranges. Instead of using one model with high computational parameters for all cases, the system adjusts model parameters (number of layers, bit depth) to match specific detection scenarios, reducing overall processing requirements while maintaining accuracy.
3Device complexity
If a single pre-trained model is used for all target sizes, then device complexity is reduced, but detection accuracy for specific target sizes deteriorates
Solution Approach 1:
The patent applies local quality by creating models with specialized characteristics tailored to specific target size ranges. Each model in the parallel system has optimized parameters and architecture suited for detecting particular size categories, ensuring high detection accuracy for specific targets while keeping individual model configurations relatively simple.
Data Source
AI summary
A target detection device including: a storage unit configured to store a plurality of pre-trained models using a deep learning network model for detecting a target from an image; a camera unit configured to acquire an image; and a target detection processing unit configured to execute parallel processing on the acquired image using the plurality of pre-trained models and detect a target based on detection results of the pre-trained models, wherein each of the pre-trained models differs from the rest of the pre-trained models in a size of a target to be extracted.