Deep learning
Great maintenance time prediction for wind turbine
There are wind turbine operation data, event code data, and maintenance time data obtained from two wind farms in Changhua Coastal Industrial Park. Then, the above data is preprocessed for filtering error information, reducing data redundancy. Moreover, apriori. algorithm is used to analyze the relationship between event code and maintenance time data. Finally, preprocessed data is used as training data to train the predictive model with CNN.
Single image dehazing via deep learning-based image restoration
Image dehazing is a challenging ill-posed problem. Existing method still has some problem about colour cast and noise. For this reason, we proposed the method that combines the deep learning with image processing. By using image processing to simplify the dehazing problem, we can more concise structure of the neural network to solve this problem with a light computation. This method can be for most application.
Retinal entropy image in deep learning to detection for diabetic retinopathy
Diabetic retinopathy (DR) is one of the microvascular complication related to diabetes mellitus and a major cause of blindness globally. Deep learning is a kind of machine learning; in modern medicine, using deep learning in fundus photography has emerged as a cost-effective method and practical for automated grading of DR. Entropy images, representing the complexity of original fundus photographs, may strengthen the contrast between DR lesions and unaffected areas and increase the detection accuracy. The research results show the entropy image quantifies the information amount of the fundus photograph and efficiently accelerates the generating of feature maps in CNN.
Efficient Intra Frame Coding Using Convolutional Neural Network in HEVC
The HEVC (High Efficiency Video Coding) Standard is a significant research topic in our lab. Here, a convolutional neural network (CNN)-based method is proposed for HEVC intra frame coding. For the CNN training, we designed and trained a fully convolutional network to learn a residual image for the input image. In intra frame coding, the proposed CNN is added following by reconstructed CTU. The reconstructed CTU (original one) is enhanced its visual quality by the predicted residual where is from the trained CNN model. Then the new reconstructed CTU is generated to replace the original one as the reference block for intra prediction of the next CTU to be encoded. After all of the CTU encoded, the reconstructed I-frame is generated by deblocking and SAO process, and it will be seen as the reference frame in encoding frame buffer.
Forecasting the approval rate of politicians based on social media
Social media is an online platform where people create, share opinions, and exchange ideas. In order to analyze the approval rate of politicians, we collect the Facebook and PTT comments. Moreover, NTUSD and deep learning tools are used to analyze the approval rate of politicians.