Related Experiment Video
Updated: Dec 20, 2025
13:19
Deep Neural Networks for Image-Based Dietary Assessment
Published on: March 13, 2021
9.8K
Design of an Always-On Image Sensor Using an Analog Lightweight Convolutional Neural Network
Jaihyuk Choi1, Sungjae Lee2, Youngdoo Son2
1Department of Semiconductor Science, Dongguk University-Seoul, Seoul 04620, Korea.
Sensors (Basel, Switzerland)
|June 4, 2020
Summary
This study introduces an always-on CMOS image sensor with an analog convolutional neural network for efficient mobile image classification. It significantly reduces power and processing time using novel analog circuits, achieving 89.33% accuracy.
Area of Science:
- Integrated Circuits
- Computer Vision
- Artificial Intelligence
Background:
- Traditional image classification in mobile applications faces challenges with high power consumption and processing time.
- Complementary Metal Oxide Semiconductor (CMOS) image sensors (CIS) are prevalent but often require significant digital processing for complex tasks like image classification.
Purpose of the Study:
- To develop an always-on CIS with an integrated analog convolutional neural network (CNN) to reduce power consumption and processing time for mobile image classification.
- To implement analog circuits for core operations like convolution and max-pooling, minimizing reliance on power-hungry digital components.
Main Methods:
- Proposed analog convolution circuits for convolution, max-pooling, and correlated double sampling, eliminating the need for operational transconductance amplifiers.
- Utilized a voltage-mode MAX circuit for analog max-pooling and a 4-bit single-slope analog-to-digital converter for data conversion after analog processing.
- Fabricated a prototype CIS using a 0.11 μm 1-poly 4-metal CIS process with a standard 4T-active pixel sensor, achieving 160x120 resolution.
Main Results:
- Achieved a 99.58% reduction in image data after analog convolution processing.
- Demonstrated an 89.33% image classification accuracy with the integrated analog CNN.
- The prototype CIS operated at a total power consumption of 1.12 mW with a 3.3 V supply voltage and a maximum frame rate of 120 frames per second.
Conclusions:
- The proposed always-on CIS with an analog CNN effectively reduces power consumption and processing time for mobile image classification tasks.
- Analog implementation of key neural network operations offers a viable path towards more efficient on-sensor processing.
- The system achieves competitive classification accuracy while maintaining low power operation, suitable for mobile and embedded applications.
