| Title |
Quantization Driven Lightweight Deep Neural Network for Image Classification |
| Authors |
최준우(Junewoo Choi) ; 이덕우(Deokwoo Lee) |
| DOI |
https://doi.org/10.5573/ieie.2024.61.11.172 |
| Keywords |
Backbone network; Lightweight; Quantization; Object detection |
| Abstract |
Recently, as high-performance deep learning models with deep and vast structures have been introduced one after another, the required computing resources are increasing. In this case, in order to expand the usability of the model, a method of lightening the weight of the model is essential, and typically, there is an autonomous vehicle that uses object detection in a field where real-time reasoning is important. In the case of the object detection model, the image classification model used as the backbone network among components is generally used. By replacing this backbone network with an efficient model, the number of parameters and required computing resources can be effectively reduced. In this paper, we propose an improved MobileNetV3 by replacing the activation function of the MobileNetV3 model and applying the quantization technique among the image classification models used as the backbone network in the object detection model. The main change of the model was that the activation function was changed to Leaky ReLU, the hyperparameter was selected as the optimized value through Bayesian optimization, and the model was lightened by performing quantization. The experimental results confirmed that these improvements contributed to maintaining the accuracy while reducing the weight of the model. |