De-raining of image based on machine learning

• Main Author: Peng, Xiaohua
• Other Authors: Jiang Xudong
• Format: Theses and Dissertations
• Language: English
• Published: 2019
• Subjects: DRNTU::Engineering::Electrical and electronic engineering
• Online Access: http://hdl.handle.net/10356/78542
• Institution: Nanyang Technological University

In recent several decades, rain removal task, especial single image de-raining has attracted pretty much attention. To address the de-raining problem, in this dissertation, we propose an improved algorithm based on deep learning frameworks which have been proved to be effective on rain removal tasks. On the one hand, inspired by RESCAN network architecture, which extends context aggregation net (CAN) with squeeze and excitation (SE) blocks and the dilated convolution to construct a novel deep learning architecture, we first propose a similar network architecture named CSECAN, instead of using RNN structure, we employ deep convolutional neural network in our model. Because CNN is intrinsically more applicable for handling images, which is also verified by the final performance comparison. On the other hand, to make the network faster and easier to train, we adapt a prior decomposition process into our model and propose our second model named as De-CSECAN. By utilizing the decomposition process, we are able to decompose the input image into low and high frequency layer, and we can decrease the mapping range effectively by inputting the high frequency layer into our model, which leads to a faster neural network. Though this model proves to be not so effective on de-raining, the idea of speeding up the training process by reducing the whole mapping range is still worth to be noticed. At the discussion part, we also analyze the reasons of our second model’s ineffectiveness. To obtain a more objective performance comparison rather than just by human eyes’ looking between different algorithms on practical cases, we construct our own real-world paired dataset termed as Constructed Paired Dataset, which consists of 9 real case raining images and their corresponding no-rain images. Finally, by conducting extensive experiments on various de-raining approaches on both synthetic and real-world dataset, we demonstrate the effectiveness of our model on performing single image rain removal.