good articles about Deep learning and its applications
1. Transferring Deep Convolutional Neural Networks for the Scene Classification of High-Resolution Remote sensing Imagery
It’s about the transfer learning method using in remote sensing field
Value: Transfer learning, using pretrained model, and using a few training samples to get good result.
target:
- Investigate how to transfer features from these successfully pre-trained CNNs for HRRS scene classification with limited training data
- Scenes classification
using many networks to try, including:
- AlexNet 2.CaffeNet 3.VGG-F 4.VGG-M 5.VGG-S 6.VGG-VD16 7.VGG-VD19 8.PlacesNet
Dataset:
- UCM
- WHU-RS
Some tricks:
- We usually only using the final layers, actually using middle layers is also effective
- Based on the trained model
- Only a few training data
- By removing the last few layers of a CNN, treat the remainder of the CNNs as a fixed feature extractor
- regard the activation vector of the fully-connected layer as the global feature representations for scenes, using other sample classifiers such as linear SVM to be the classifier
- generate dense CNN activations from the last or penultimate convolutional layer with multiple scales, aggregate the obtained features into a global representation via the conventional feature coding scheme: such as BOW, Fisher encoding….
- Compare the different feature coding methods, effect of different convolutional layers, and compare with low-level features
- Compared all the result, and proved then corresponding method is state-of -the-art
Others:
- To intuitively understand the CNN activations, visualize the representation of each layers by inverting them into reconstruction images with the technique proposed in the paper called: Understanding Deep Image Representations by inverting them.
- Alex,Caffe,VGG-F and PlacesNet have the similar structure.
2. Classification and Segmentation of Satellite Orthoimagery using convolutional neural networks
Best point is about using multiple CNN layers, and pretrained CNN filters with K-meaning
Target
- Classify different landscapes, especially buildings
Some Tricks:
- Using near infrared bands, totally 14 bands, some of them overlapped
- Digital surface model(DSM), increases classification accuracy by providing height information that can help distinguish between similar looking categories.
- Using a GUI to manually labeling, pre-processing using the cluster algorithm called simple linear iterative clustering
- pixel to be classified is located at the center
- filter of CNN pretrained by unsupervised methods such as K-meaning
- Using multiple CNNs, concatenated as in put for the fully-connected layer. Two different methods
- Different types of CNN structure between each level
- Different input, same CNN structure
- Post-Processing
- Classification Average: can get rid of salt-and-pepper
- Region merging: merge the result with original segmentation result, set a threshold, two adjacent segments were merged if the average classification accuracy of all the pixels within each segment were over the threshold
- Prove and compare many aspacts
- how to choose the best architecture: Window size, filter size, filter number, pooling dimension
- best number of CNN
- Influence of the number of CNN in parallel
- different CNN numbers
- different parameter settings
- Good result come out when combine:
- aforementioned best trucure
- Average combine
- Region Merging
Others:
- Very good generalization of CNN using in Remote sensing