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

抄録

Anticipating the rise of the Internet of Things (IoT) era, we have proposed an object detection framework that employs a CMOS image sensor with binary feature extraction to reduce power requirements. Initially, we presented a lightweight deep neural network for the feature data based on the YOLOv7, comparable to the YOLOv7-tiny in the number of parameters and FLOPs, but it enhances large object recognition accuracy (APL50) by 6.6%. Moreover, our approach achieves a 48.8% reduction of GPU power consumption compared to the YOLOv7. Additionally, we introduce an on-chip signal processing method for the binary feature data. The proposed method achieves a compression rate of 64.1% and increases GPU power consumption by only 14.9% during the decoding process preceding object detection. Moreover, the size of 1-bit feature data is reduced by 96.0%, and object recognition accuracy is improved by 4.0% relative to 1-bit RGB color images.