| Title |
A Study on Hardware-friendly Quantization for Object Detection Networks using Integer-only Arithmetic Operations |
| Authors |
서유권(Yugwon Seo) ; 강진구(Jin-Ku Kang) ; 김용우(Yongwoo Kim) |
| DOI |
https://doi.org/10.5573/ieie.2025.62.4.75 |
| Keywords |
Deep-learning; Hardware-friendly; Integer-only arithmetic; Object detection; Quantization |
| Abstract |
With recent advancements in hardware technology, deep learning has been widely applied in various fields, including computer vision. In particular, CNNs are utilized in the field of computer vision, and adding more convolutional layers results in improved performance. However, this requires significant memory and computational resources, making it difficult to apply CNNs in embedded environments. One solution to this problem is quantization, a technique that converts 32-bit floating-point data into lower-bit integer data. Among quantization methods, one method employs techniques such as representing the scale factor as a power of two to perform hardware-friendly quantization. However, this method fuses the batch normalization layer before performing quantization-aware training, which prevents the model from leveraging the benefits of the batch normalization layer during retraining. In this paper, we propose a hardware-friendly quantization-aware training method that modifies the batch normalization layer fusion process to retain the advantages of batch normalization layer during training. The proposed method was applied to the YOLOv7-tiny object detection model, and experimental results demonstrated up to a 1.5%p improvement in mAP(0.5) compared to existing method. Additionally, when the quantized model was converted into a network using only integer operations, further testing showed up to a 2.5%p improvement in mAP(0.5). |