Chip Design for a Real-Time Artificial Intelligence Vi Deo Image Detection with High Dynamic Range Technology

This project aims to the chip design of a real-time artificial intelligence video detection with high dynamic range (HDR) technology. The proposed chip includes HDR converter chip and real-time AI video detection chip. The proposed HDR chip can converter the low dynamic range (LDR) video stream to HDR stream from any cameras. The proposed AI chip, called Multi-View Recursive Convolutional Neural Network (MV-ReCNN), can improve the accuracy of the well-known neural network, such as VGG16, VGG19, DenseNet, …etc. The HDR video will input to the proposed real-time MV-ReCNN chip to achieve high accuracy of detection. On the other hand, the HDR algorithm converts the images to the LDR images. These LDR images and the original images can input the neural network as the training data to improve the detection accuracy.This project is planned for three years. In the first year of this project, the algorithm of HDR and MV-ReCNN, included the network and database established, will be completed. In the second year, the prototype of the FPGA-based system and ASIC, includes HDR and MV-ReCNN chips, are implemented. In the final year, the AI-HDR chip will be implemented. The proposed chip owns HDR conversion and real-time AI detection, and all the chips of this project are designed, implemented, evaluated, integrated based on the TSMC 90/45 nm CMOS technology.

Project ID:PB10907-2539
External Project ID:MOST109-2221-E182-050