ECS795P Deep Learning and Computer Vision, 2023
Course Work 1: Image Super-resolution Using Deep Learning
代写深度学习作业 Coding: to add your code blocks in the required sections; (40% of this CW)Report: to complete the questions in report; (30% of this CW)
Introduction
Aim: To obtain practical knowledge and hands-on understanding of the concepts of image super-resolution, deep learning using convolutional neural networks (CNN) and peak signal-to-noise ratio (PSNR).
Start: Download CW1_ECS795P.zip from the course website at
http://www.eecs.qmul.ac.uk/~sgg/_ECS795P_/.
Tasks: three subtasks are involved:
- Coding: to add your code blocks in the required sections; (40% of this CW)
- Report: to complete the questions in report; (30% of this CW)
- Online assessment: to answer one question and to conduct one exercise,which are randomly selected from below. It will be carried out during the lab demo session in WK10; (30% of this CW)
Platform: Python + PyTorch
- Understanding image super-resolution
Objective: To become familiar with the image super-resolution problem setting.
Questions: 代写深度学习作业
- What is the concept of image resolution?
- What is the gray-scale or single-channel image super-resolution?
- What is the Ground Truth image?
- How to measure the quality of the output high-resolution images?
Exercises:
- To read the image named butterfly_GT.bmp
- To show the image resolution of this image
- To convert the image from the RGB colour space into the gray-scale space
(Tip: use imageio package to read image)
- To downsample the current image by 3 times
- To upsample the current image by 3 times with interpolation algorithm2. Understanding deep learning by convolutional neural network
Objective: To understand the principles of deep convolutional neural network.
Questions: 代写深度学习作业
- What are the parameters of a CNN?
- What is the target of CNN model training?
- What is the difference between the training and testing stage of a CNN?
- What is the feature map?
- How to conduct the convolutional operation?
Exercises:
- To load the pre-trained model named model.pth
- To set and show the weights of the first convolutional layer
- To set the channel number of the input
- To set the filter number
- To set the filter size
- To set the padding
- To show the weight of the 1st filter in command window
- To show the bias of the 10th filter in command window
- To set and show the weights of the second convolutional layer
- To set the channel number of the input
- To set the filter number
- To set the filter size
- To set the padding
- To show the weight of the 5th filter in command window 代写深度学习作业
- To show the bias of the 6th filter in command window
- To set show the weights of the third convolutional layer
- To set the channel number of the input
- To set the filter number
- To set the filter size
- To set the padding
- To show the weight of the 1st filter in command window
- To show the bias of the 1st filter in command window
- To perform 2-d convolutional operation on a 2-d matrix with a given filter
(Tip: conv2d and relu are PyTorch build-in functions)3. Image super-resolution using deep convolutional network
Objective: To perform image super-resolution with deep convolutional neural network and evaluate the performance.
Questions: 代写深度学习作业
- How to use a trained SRCNN to perform image super-resolution (testing stage)?
- What are the input and the output of the SRCNN?
- How to conduct qualitative and quantitative comparison between two different SR networks?
- What is the typical numerical measure metric for quantitative analysis?
- What is the maximum power of imaging signal (i.e., pixel) and noise signal? e.g., image of uint8 type?
Exercises:
- To get and show the Ground Truth image
- To get and show the low-resolution image (downsampled) 代写深度学习作业
- To feed the input image into the SRCNN
- To get and show the output high-resolution image by SRCNN
- To compute the PSNR of the high-resolution image against the Ground Truth image (Tip: use the python module: skimage.metrics.peak_signal_noise_ratio)
- To get the high-resolution image by interpolation algorithm (baseline result)
- To compute the PSNR of the baseline result against the Ground Truth
- To compare the results of the two methods (baseline and SRCNN) in terms of PSNR
