Layer-Wise Weight Quantization of Object Detection Models Using Quantized Features Extracted from a CMOS Image Sensor

This paper investigates weight quantization and pruning for object detection models trained on aggressively quantized feature image extracted from CMOS image sensors. We first coarsely quantize the DNN model to demonstrate that aggressively quantized feature images exhibit high robustness to model weight quantization. Based on this finding, we propose a layer-wise fine weight quantization method suitable for the quantized feature images and a pruning technique. Compared with the YOLOv7-tiny trained on RGB color images, the proposed model trained on the feature images reduced the input data size by approximately 97.66%, the model size by 94.11%, and the number of multiply-accumulate operations (MACs) by 70.70%. Meanwhile, the degradation in detection accuracy for large objects, mAPL50, was limited to 1.1%.