Compact Convolutional Neural Network for Mobile Face Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks are impractical for low-resource processing devices due to their large size, making it difficult to run them on mobile devices for image processing and classification tasks.
Innovation Solution
A compact convolutional neural network structure is designed, comprising a preliminary layer group, intermediate layer groups, and a final layer group, with specific layers such as normalization, padding, convolution, activation, and downsampling layers, optimized for reduced resource usage, including reduced precision filters from 32 bits to 5 bits for storage, allowing efficient operation on mobile devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional convolutional neural networks are used for image processing, then accuracy and processing capability are improved, but device size and resource requirements increase making them impractical for mobile devices
Solution Approach 1:
The network is divided into three distinct layer groups (preliminary, intermediate, and final) with specific functional segments. Each group contains specialized layers (e.g., normalization, padding, convolution, activation, downsampling) that work together to process images while maintaining compactness. This segmentation allows the network to achieve high accuracy through functional decomposition while keeping each segment small and efficient for mobile deployment.
Solution Approach 2:
The patent applies parameter changes by reducing filter precision from 32 bits to 5 bits for storage, significantly decreasing memory requirements. Additionally, the network uses specific dimensional parameters (e.g., 256 input channels, 256 output channels in intermediate layers) to optimize the balance between computational accuracy and resource consumption, enabling high-accuracy image processing on resource-constrained mobile devices.
2Device complexity
If the number of parameters is reduced to enable mobile deployment, then device resource usage is improved, but network accuracy may deteriorate
Solution Approach 1:
The patent changes the precision parameter of filters from 32 bits to 5 bits for storage, reducing the number of parameters by a factor of 7.87 while maintaining face recognition accuracy. This parameter change enables the network to be deployed on mobile devices with limited memory resources without sacrificing the accuracy required for reliable face recognition and image classification tasks.
Solution Approach 2:
By segmenting the network into three functional layer groups with specific intermediate layers (squeeze, expand, concatenation, combination), the patent maintains computational effectiveness while reducing overall parameter count. Each segment is optimized for specific processing tasks, allowing the reduced-parameter network to achieve comparable accuracy to larger conventional networks.
3Quantity of substance
If filter precision is reduced from 32 bits to 5 bits, then storage size is reduced by 51.5, but computational precision may be compromised
Solution Approach 1:
The patent changes the storage precision parameter of filters from 32 bits to 5 bits, reducing storage size by a factor of 51.5. This parameter change is specifically applied to filter weights while maintaining network functionality, enabling compact storage for mobile devices. The reduction in precision is managed through the overall network architecture design that compensates for lower filter precision through optimized layer configurations and processing techniques.
Data Source
AI summary
A compact convolutional neural network may include a preliminary layer group, one or more intermediate layer groups, a final layer group, and/or other layers/layer groups. The preliminary layer group may include an input layer, a first preliminary normalization layer, a preliminary padding layer, a preliminary convolution layer, a preliminary activation layer, a second preliminary normalization layer, and a preliminary downsampling layer. One or more intermediate layer groups may include an intermediate squeeze layer, a first intermediate normalization layer, an intermediate padding layer, a first intermediate expand layer, a second intermediate expand layer, an intermediate concatenation layer, a second intermediate normalization layer, an intermediate activation layer, and an intermediate combination layer. The final layer group may include a final dropout layer, a final convolution layer, a final activation layer, a first final normalization layer, a final downsampling layer, a final flatten layer, and a second final normalization layer.