In Pixel Ultrasonic Processing for Data Reduction and Edge AI

This effort focuses on reducing theĀdata volume generated by imagers using novel pixel-level image processing in the analog domain using GHz ultrasonic pulses. The GHz ultrasonic wave pulses compute Sonic Fast Fourier Transforms (SonicFFTs)ĀĀby using diffraction and automatic summations of different wavelets at the receive plane,ĀĀenabling massively parallel computation by propagating waves from CMOS integrated piezoelectric transducers. SonicFFTs will also be used to implement ultrafast and low power ultrasonic convolutions by taking the Inverse FFT of the product of two FFTs.ĀĀSonicConv and ultrasonic FFT (SonicFFT) will lead to tremendous data compression, on which the backend NN can be trained. This effort will develop analytical and simulation models to develop AI algorithms that will result in greatly reduced SWAP. Hardware components of SonicFFT will be implemented on CMOS with integrated photodiodes to capture the images, and generating SonicFFTs using piezoelectric transducers.