CNN Face Retrieval via FPGA Bit Width Configuration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional video monitoring systems face challenges in efficiently extracting useful information from massive images and videos due to high computation requirements of convolutional neural networks, leading to low efficiency in face image retrieval and real-time detection.
Innovation Solution
A face image retrieval method and system that configures a convolutional neural network with convolution calculation configuration information, including data bit width values for each layer, and utilizes a Field-Programmable Gate Array System on Chip (FPGA SoC) for efficient data processing, reducing network bandwidth and power consumption, and enabling real-time face detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If convolutional neural networks are used for face recognition, then recognition accuracy is improved, but computation time increases leading to low retrieval efficiency
Solution Approach 1:
The patent segments the convolutional neural network into multiple independent calculation stages (first convolutional calculation, second convolutional calculation, etc.), where each stage processes data independently. This segmentation allows parallel processing of different data streams, reducing overall computation time while maintaining recognition accuracy through multi-stage feature extraction and matching.
2Measurement precision
If high-definition monitoring devices are deployed, then image quality is improved, but data storage and transmission requirements increase
Solution Approach 1:
The patent extracts only the essential face feature information from high-definition images through convolutional neural network processing, rather than storing or transmitting the complete high-resolution images. The CNN extracts key facial features and characteristics, converting large-volume image data into compact feature vectors that retain recognition accuracy while significantly reducing data storage and transmission requirements.
Data Source
AI summary
A face image retrieval method includes: obtaining to-be-retrieved face information corresponding to a to-be-retrieved image by a convolutional neural network, the convolutional neural network being configured with corresponding convolution calculation configuration information by a processor, the convolutional neural network including at least one convolutional layer, the convolution calculation configuration information including a data bit width value corresponding to each convolutional layer in the convolutional neural network, and the to-be-retrieved image including at least one face region; searching a database for matched preset face image information that matches the to-be-retrieved face information, the database storing at least one piece of preset face image information; and outputting the preset face image information that matches the to-be-retrieved face information.