Lightweight Object Detection Model for a CMOS Image Sensor with Binary Feature Extraction

For the coming IoT age, we propose an object detection system using a CMOS image sensor capable of extracting binary feature data in order to reduce the power consumption of recognition systems. First, a lightweight deep neural network (DNN) for feature data is verified based on YOLOv7. The presented model is comparable to the YOLOv7-tiny in the number of parameters and FLOPs, but improves the object recognition accuracy of large objects AP
L
50 by 6.6%. Additionally, the presented model for the feature data reduces GPU power consumption by 39.2% compared to the original YOLOv7 that processes RGB color images. Secondly, an on-chip signal processing technique for the CMOS image sensor is proposed to extract binary feature data. The simulation results demonstrate that the size of 1-bit feature data is reduced by 96.0% at the cost of a image classification accuracy degradation of 8.3% compared to the 1-bit RGB color image.