MiniNet:密集擠壓之深度可分離卷積於圖像分類 MiniNet: Dense Squeeze with Depthwise Separable Convolutions for Image Classification

Kao, Fan-Yi
Journal Title
Journal ISSN
Volume Title
近年來,人工智慧的發展蒸蒸日上,自卷積神經網路被提出之後,深度學習開始蓬勃發展,研究學者們紛紛提出更為優化與創新的技術,相較於其它科學領域,深度學習領域的研究採完全開放的方式進行,Google團隊提出TensorFlow開放原始碼函式庫,並在TensorFlow核心庫中支援高階深度學習框架的Keras,幫助開發者在Keras中建立及訓練深度學習模型。未來人工智慧的應用將無所不在,為普及自動駕駛、無人商店、智慧城市等應用,如何在有限的硬體設備中,提供一個運算快速且低計算成本的神經網路模型已成為一個很重要的研究議題。 本論文基於MobileNet架構,加入密集連接技術與擠壓式的SENet模型,提出一個密集擠壓之深度可分離卷積架構,並將此模型命名為MiniNet。本論文在實驗環境中,使用Keras進行MiniNet的建立與訓練,在五種不同的資料集中,與三個現有的卷積神經網路架構進行比較,實驗結果顯示,本論文提出之MiniNet架構能夠明顯地使用更少的計算參數量並有效地縮短訓練時間,尤其在資料集之種類與資料量較少時,本論文提出之MiniNet架構更能優於現有架構達到最高的準確率。
Artificial intelligence (AI) has been developed vigorously in recent years. Deep learning has made a breakthrough since the convolutional neural network was proposed. Researchers have proposed various improved and innovative techniques. Compared with other research fields, researches in deep learning are conducted with an open-source environment completely. The Google Brain team developed the open-source library TensorFlow, which supports the Keras functional API and helps developers to build and train deep learning models. AI applications will be ubiquitous in the future, such as self-driving cars, unmanned shops, and smart city applications. How to decrease computations and shorten calculation time is a vital research issue. In the thesis, the dense squeeze with depthwise separable convolutions, viz MiniNet is proposed and based on MobileNet’s architecture. Not only the dense connected technique but also the Squeeze-and-Excitation concept are integrated into MiniNet. To build and train the proposed MiniNet, experiments in the thesis are implemented with Keras. Three existing models are compared to the MiniNet with five different datasets. The experimental results showed that the MiniNet significantly reduces number of parameters and shortens training time efficiently, achieves the highest recognition accuracy when the dataset is small especially.
深度學習, 卷積神經網路, 深度可分離卷積, 密集連接, Deep Learning, Convolutional Neural Network, Depthwise Separable Convolution, Densely Connected Convolutional Networks