전체 글6 [안생강 논문리뷰 03] ECA-Net: Efficient Channel Attention for Deep Convolutional Neural Networks ( 2021.04.19 updated ) Paper Link / GitHub Link (2020) 1. Introduction SE-Net : squeeze-and-excitation, learns channel attention - input > GAP for each channel independently > two FC layers with non-linearity followed by a sigmoid function ( two FC layers are designed to capture non-linear cross-channel interaction, which involve dimensionality reduction for controlling model complexity ) - rese.. 2021. 4. 19. [안생강 논문리뷰 02] CBAM: Convolutional Block Attention Module ( 2021.04.18 updated ) Paper Link / GitHub Link / Blog (2018) 1. Introduction Three important factors to enchance performance of CNNs - depth : LeNet, VGGNet, ResNet, etc. - width : GoogLeNet, etc. - cardinality : Xception, ResNeXt, etc. Attention mechanism - tells where to focus and improves the representation of interests - focusing on important features and suppressing unncessary ones In this.. 2021. 4. 18. [안생강 논문리뷰 01] ChestNet: A Deep Neural Network for Classification of Thoracic Diseases on Chest Radiography ( 2021.04.17 updated ) Paper Link (2018) 1. Introduction Computer-aided diagnosis of thorax disease on chest radiography - due to chest radiography's low-cost and easy access nature, it is one of the most common types of radiology examinations for the diagnosis of thorax disease Challenging tasks - complexity and diversity of thorax disease and the limited quality of chest radiographs - few publ.. 2021. 4. 17. Paper Review LIST ( 2021.04.18 updated ) Chest Radiography ChestNet: A Deep Neural Network for Classification of Thoracic Diseases on Chest Radiography Attention Mechanism CBAM: Convolutional Block Attention Module ECA-Net: Efficient Channel Attention for Deep Convolutional Neural Networks 2021. 4. 17. 1x1 convolution의 역할과 Bottleneck 📌 1x1 convolution의 역할을 요약하자면, - dimension reduction - 각 픽셀별 위치정보를 해치지 않은 채 필터들의 조합에 따른 정보 분석 convolution 연산을 통해 주변 레이어와의 spatial 정보까지 고려한 패턴을 보고 특징을 추출한다. ( 예를 들어 3x3 convolution을 하면 3x3 공간의 특징을 하나의 값으로 추출한다 ) 1x1 convolution은? "하나의 픽셀"만 고려하므로 ( channel 정보만 O, spatial 정보 X ) channel 방향으로만 계산해서 1개의 숫자를 출력해 dimension reduction의 역할을 하게 된다. 여기서 언제나 차원이 축소되는 것은 아니고, filter 개수에 따라 output dimension이 달라.. 2021. 4. 14. 이전 1 2 다음
