[컴퓨터비전] CNN

What is CNN

• Convolutional Neural Networks
• CNNs stack CONV, POOL, FC layers.

Convolution layer (CONV)

• CONV uses filters that perform convolution opereations as it i scanning the input image with respect to its dimensions.
• The result of convolution must be two dimensional.
• Padding 통해서 input, output 사이즈를 맞춘다.

Pooling layer (POOL)

• POOL is a downsampling operation.

Weight initialization

• sigmoid and tanh activation function -> Xavier 적합
• ReLU -> He 적합
• 대부분 He initialization 사용

Backpropagation

• Feed the error backwards through the network
• Ultimately adjut the weights of all the connections of the network

Techniques to reduce overfitting

• Increase training data
• Reduce model complexity
• Early stopping during the training phase
• Add regularization
• Use dropout for neural networks

CNN Model Architectures

https://velog.io/@ssoyeong/%EB%94%A5%EB%9F%AC%EB%8B%9D-CNN-Architectures

DenseNet

• 하나의 레이어에만 전달하는 것이 아니라, 여러 개 dense하게.
• Strengthen feature propagation

FPN (Feature Pyramid Network)

• Detecting objects at different scales
• 서로 다른 해상도의 feature maps를 쌓아올린 형태
• Bottom-up, top-down 방식 두 가지를 적용해서 더 좋은 feature를 생성한다.

Squeeze-Excitation(SE) block

• explicitly modelling the interdependencies between the channels of its convolutional features.
• 중요한 channel을 찾는다. Channel attention을 고려해야 한다.
• Squeeze: Global information embedding, 각 feature map에 대한 전체 정보 요약
• Excitation: Adaptive Feature Recalibration, 각 feature map의 중요도를 조정

장점

• The SE block can be integrated into standard architectures.
• 파라미터의 증가량에 비해 모델 성능 향상도가 매우 크다.