Binary CNN Feature Extraction for Resource-Constrained Object Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep convolutional neural networks (CNNs) for object recognition require significant computational resources and memory, making them unsuitable for many applications, particularly in mobile devices due to their heavy burden on devices with hundreds of millions of parameters and complex feed-forward computations.
Innovation Solution
Generating binary convolutional neural network features using a subset of response maps from a full CNN, which are then fed into a strong classifier to reduce memory and computational requirements while maintaining accurate object recognition results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep convolutional neural networks are used for object recognition, then recognition accuracy is improved, but memory usage and computational resources increase significantly
Solution Approach 1:
The patent extracts and eliminates unnecessary components from the deep CNN architecture. Specifically, it removes fully connected layers and retains only the convolutional feature extraction layers, thereby reducing memory usage from 232 MB to 14 MB while preserving the essential feature extraction capability needed for accurate object recognition
Solution Approach 2:
The patent segments the deep CNN into functional modules: input layer, convolutional layers for feature extraction, and a separate classification layer. This segmentation allows the memory-intensive fully connected layers to be eliminated while maintaining the feature extraction capability of the convolutional layers, resolving the contradiction between accuracy and memory usage
2Measurement precision
If deep convolutional neural networks are used for object recognition, then recognition accuracy is improved, but computational complexity increases significantly
Solution Approach 1:
The patent extracts and removes the computationally intensive fully connected layers from the deep CNN architecture, retaining only the convolutional layers for feature extraction. This reduces computational complexity by eliminating the expensive matrix multiplications in fully connected layers while preserving accuracy through the use of convolutional feature maps
Solution Approach 2:
The patent uses binary neural features as simplified copies of the full CNN feature representations. These binary features are generated by thresholding the continuous feature maps from convolutional layers, creating a compressed representation that reduces computational complexity while maintaining the essential discriminative information for accurate object recognition
3Quantity of substance
If binary features are used instead of continuous features, then memory usage is reduced, but recognition accuracy may deteriorate
Solution Approach 1:
The patent changes the parameter representation from continuous floating-point values to binary values (0 or 1). This parameter transformation reduces memory usage by approximately 16 times while maintaining recognition accuracy because the binary features are derived from the most discriminative regions of the continuous feature maps through thresholding operations
Data Source
AI summary
Techniques related to implementing convolutional neural networks for object recognition are discussed. Such techniques may include generating a set of binary neural features via convolutional neural network layers based on input image data and applying a strong classifier to the set of binary neural features to generate an object label for the input image data.