Low Power Shallow Hash Network for Edge Image Retrieval
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing content-based image retrieval (CBIR) systems face challenges in efficiently retrieving similar images due to high computational requirements and power consumption, especially on portable devices, as they often rely on deep neural networks that are impractical for edge devices and have sub-optimal performance with large datasets.
Innovation Solution
The implementation of a low power supervised semantic-preserving shallow hashing (LPSSSH) technique using a two-pass learning method, where a deeper network generates hash codes for training, and a shallow neural network is trained with fewer parameters to achieve high accuracy with reduced compute bandwidth and power consumption, incorporating adaptive bitwise thresholding and early stopping for efficient training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep neural networks are used for image retrieval, then accuracy is improved, but power consumption and computational requirements increase
Solution Approach 1:
The patent segments the neural network into two parts: a deep network used only during training to generate hash codes, and a shallow network deployed for actual inference. This segmentation allows the system to achieve deep network accuracy during training while using a lightweight shallow network during operation, significantly reducing power consumption on portable devices.
Solution Approach 2:
The patent changes the parameter configuration by training a shallow network with hash codes generated by a deep network as supervision signals. This parameter change enables the shallow network to learn optimal hash code generation without requiring the computational resources of a deep network during inference, resolving the contradiction between accuracy and power consumption.
2Measurement precision
If deep neural networks are used for image retrieval, then retrieval quality is improved, but device complexity increases
Solution Approach 1:
The patent segments the system into training phase (using deep network) and inference phase (using shallow network). The deep network's complexity is confined to the training phase, while the deployed shallow network has minimal complexity, making it suitable for portable devices while maintaining retrieval quality through the two-pass learning approach.
Solution Approach 2:
The patent introduces hash codes generated by the deep network as an intermediary supervision signal for training the shallow network. This intermediary allows the shallow network to learn from the deep network's capabilities without directly incorporating its complexity, enabling high retrieval quality with low device complexity.
3Use of energy by moving object
If shallow neural networks are used, then power consumption is reduced, but retrieval accuracy decreases
Solution Approach 1:
The patent performs preliminary action by using a deep network to generate high-quality hash codes during the training phase. These pre-generated hash codes serve as supervision signals that guide the shallow network's training, enabling the shallow network to achieve high retrieval accuracy without requiring the computational power of a deep network during actual use.
Solution Approach 2:
The patent creates a copy of the deep network's hash code generation capability through supervision. Instead of deploying the actual deep network, the shallow network learns to replicate its hash code generation function using the deep network's outputs as training targets, achieving similar accuracy with much lower power consumption.
4Speed
If traditional hashing methods are used, then computational speed is improved, but semantic preservation capability worsens
Solution Approach 1:
The patent replaces traditional mechanical hashing methods with a neural network-based approach. The shallow neural network learns semantic features and generates hash codes that preserve semantic information, unlike traditional hashing methods that rely on fixed mathematical transformations. This substitution maintains computational speed while significantly improving semantic preservation capability.
Data Source
AI summary
A semiconductor package apparatus may include technology to provide an image to a low power shallow hash network, generate a hash code from the low power shallow hash network, and identify one or more similar images based on the hash code. Other embodiments are disclosed and claimed.