Shallow Hash Neural Network for Fast Image Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image recognition technologies using convolutional neural networks (CNN) are inefficient due to the need to perform recognition on multiple models and aggregate class probability vectors, leading to long recognition times and high calculation complexity, especially when retrieving similar images with high-dimensional feature vectors.
Innovation Solution
The implementation of a shallow hash neural network (SHNN) model, which aggregates multiple hash neural network (HNN) models to reduce recognition time while maintaining accuracy, by training the SHNN model using aggregated recognition results from multiple HNN models and converting feature vectors into binary hash codes for efficient similarity calculations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If multiple CNN models are used for image recognition to improve accuracy, then recognition accuracy is improved, but recognition time increases
Solution Approach 1:
The patent applies preliminary action by pre-training multiple CNN models offline before actual image recognition tasks. The models are trained in advance on large datasets, and their weights are aggregated to create a composite model. During runtime, only the aggregated model needs to be used, significantly reducing recognition time while maintaining the accuracy benefits of multiple models.
Solution Approach 2:
The patent merges multiple CNN models by aggregating their weight parameters into a single composite model. The aggregation process combines weights from multiple pre-trained models (e.g., VGG16, VGG19, ResNet) to create an ensemble model that captures the strengths of each individual model, achieving improved accuracy without the computational overhead of running multiple models sequentially.
2Reliability
If multiple CNN models are used for image recognition to improve accuracy, then recognition accuracy is improved, but calculation complexity increases
Solution Approach 1:
The patent performs the computationally intensive model training and weight aggregation in advance during an offline phase. This preliminary action transfers the calculation complexity from the online recognition phase to the offline preparation phase, making the actual image recognition process computationally efficient while still benefiting from the combined expertise of multiple models.
Solution Approach 2:
The patent creates a simplified copy or representation of multiple models by aggregating their weights into a single composite model. Instead of maintaining and executing multiple separate model structures, the system uses a single model that encapsulates the collective knowledge of multiple pre-trained models, reducing computational complexity while preserving accuracy benefits.
3Measurement precision
If high-dimensional feature vectors are used for image retrieval to improve accuracy, then retrieval accuracy is improved, but retrieval time increases
Solution Approach 1:
The patent applies parameter changes by transforming the feature representation from high-dimensional continuous vectors to compact binary hash codes. This parameter transformation reduces the dimensionality and changes the data type, enabling faster similarity computations while maintaining retrieval accuracy through carefully designed hashing functions that preserve semantic relationships.
Solution Approach 2:
The patent substitutes the traditional mechanical computation of high-dimensional vector similarities with a more efficient binary code-based retrieval system. By replacing continuous mathematical operations with discrete binary comparisons, the system achieves faster retrieval speeds while maintaining accuracy through optimized hashing strategies that preserve the semantic structure of the data.
4Measurement precision
If high-dimensional feature vectors are stored for image retrieval to improve accuracy, then retrieval accuracy is improved, but storage requirements increase
Solution Approach 1:
The patent changes the storage parameter from high-dimensional floating-point vectors to compact binary hash codes. This parameter transformation dramatically reduces the storage space required per image while maintaining retrieval accuracy, as the binary codes preserve the essential semantic information needed for similarity search in a much more space-efficient format.
Data Source
AI summary
In some examples, processing circuitry obtains a shallow hash neural network (SHNN) model that has been trained from a HNN model and based on a set of SHNN training images that aggregates image recognition results from at least two reference hash neural network (HNN) models. Further, the processing circuitry performs an image recognition on the image according to the SHNN model, to obtain an image class vector in an image class space. The image class vector includes probability values of respective image classes in the image class space. A probability value of an image class in the image class space is a combination of intermediate probability values of the image class that are resulted from the at least two reference HNN models. Further, the processing circuitry determines one of the image classes for the image according to the probability values of the respective image classes in the image class space.