自從2019年COVID-19疫情爆發至今,全球醫療及經濟受到重大挑戰,為了減少醫療負擔,用於COVID-19檢測的深度學習研究備受關注。在本文中,我們使用胸部X射線數據集,比較限制對比度自適應直方圖均衡化(CLAHE)在數種基於注意力(Attention)的卷積神經網路上的影響。 本研究以深度學習之卷積神經網路進行COVID-19新冠肺炎偵測,並使用COVID-19、正常及病毒性肺炎三種分類的數據集,去評估當使用CLAHE之後,對深度學習分類效能之影響為何。本研究使用VGG16、VGG19、ResNet、NASNet、Xception等五種卷積神經網路架構,搭配ReLU、Mish、Swish三種激勵函數及Adam、Nadam、RMSProp三種優化器,並在全連接層前加入注意力層增加效能,以共45種組合比較使用CLAHE對分類檢測造成的差異,找出最適合運用CLAHE的演算法組合。
Since the outbreak of the COVID-19 epidemic in 2009, the global medical economy has faced major challenges. We are concerned about reducing medical loads and using them for COVID-19 detection and deep learning research. We will use the chest X-ray images data set in this article to compare the impact of Contrast Limited Adaptive Histogram Equalization (CLAHE) on several kinds of attention-based neural networks. In this thesis, COVID-19 detection was carried out with the convolutional neural network of deep learning, and the effect of deep learning classification was assessed using data sets containing three categories: COVID-19, normal and viral pneumonia and evaluates the impact of CLAHE on the classification effect of the data. This study employs VGG16, VGG19, ResNet, NASNet, and Xception five convolutional neural network architecture, using ReLU, Mish, Swish three activation functions and Adam, Nadam, RMSProp three optimizer, and in front of the full connection layer to add attention layer to increase performance, with a total of 45 combinations to compare the differences caused by CLAHE approach in classification detection, to find out the most suitable use of CLAHE in deep learning algorithm combination.