Fig. 4: Spectral convolutional neural network (SCNN) chip implemented by utilizing pigments as spectral filters and achieving mass production on a 12-inch wafer.
From: Spectral convolutional neural network chip for in-sensor edge computing of incoherent natural light
a The fabricated SCNN chips on a 12-inch wafer by lithography. b A tiny camera equipped with the SCNN chip. It can achieve in-sensor edge computing and spectral sensing. The size of the SCNN chip is only about \(3\times 3.5{{mm}}^{2}\) and the size of the whole camera is about \(6.5\times 7m{m}^{2}\) c A microscope image of the fabricated SCNN chip. It has 9 convolutional kernels of size \(1\times 1\) and stride \(1\times 1\). A super-pixel contains 9 image sensor pixels. d The focused ion beam-scanning electron microscope (FIB-SEM) image of the SCNN chip. One image sensor pixel is covered by a pigment-based spectral filter and a micro-lens. The fabrication process is completely standard semiconductor lithography process. e We place the thyroid pathological sections right above the lens without a microscope. f The sections and the corresponding feature maps outputted by optical convolutional layer (OCL). g The face anti-spoofing results of the pigment-based SCNN. ENL: Electrical convolutional layer.
