关键词: 读出电路, 高速, 低功耗, 卷积神经网络, 人脸检测

Abstract: In CMOS image sensor used for face detection, both frame rate and power consumption are key metrics. To address the issues of low frame rate and high power consumption in smart face detection image sensors, a high-speed, low-power readout circuit is proposed. The convolutional neural network (CNN) is used for face detection, which compresses the output data of the original image by 66.6% and reduces the number of analog-to-digital conversion. The CNN used can achieve 3×3 convolutions, rectified linear units, 2×2 max pooling, and 1×1 fully connected layers. The proposed readout circuit supports configurable timing, making readout circuit to applicable in both grayscale image readout mode and CNN mode through different timing configurations. The proposed readout circuit is simulated in a 55-nm CMOS process, with a supply voltage of 1.2 V. In the CNN mode, the output frame rate reaches 603 frame/s, and the total power consumption of the readout circuit is 137 μW. The accuracy for face detection achieves 98.3% in the Labeled Faces in the Wild dataset.

Key words: readout circuit, high speed, low power, convolutional neural network, face detection