Inclusive & Assistive Products
ISDN2001/2002: Second Year Design Project
Comparing Existing Method on Image Processing and the improvement of the Quality of Image
Samuel
Objectives
Our product involves using a small portable device to take photos and transferring them to another digital photo frame, but the problem is that the photos may not be clear. Although there are existing image processing algorithms to solve the issue, which one would be the best for our product still remains unclear. This project will compare the existing algorithms on improving the image qualities, and find the optimised method.
Parameters
These are the different catigoris of image processing algorithms that are used in the test
Denoising
Denoising works by blurring. It takes pixel within a certain area and give them a value similar value
Brightness
The process increase the brightness value of pixels by a certain amount
Saturation
The process is done by adjusting the color value of each pixels and increase the contrast value
Comparing Method
Image quality is something that could be subjective. How should we compare it? Luckily, there is a metircs in Matlab called PIQE (Perception based Image Quality Evaluator), which estimates block-wise distortion and measures the local variance of perceptibly distorted blocks to compute the quality score.
Test Images
To make the images more applicable to our use case, all these photos are taken by the capturing device and set to the same resolution.
Testing Procedures
Step 1
Set up the python environment by importing libraries in the code. Then, read the images in the folder and change their resolution into 1200 x 1600.
Step 2
The image will be then pass into the processing algorithm
Step 3
Average all the results for each algorithms and do a comparison
Results & Analysis
By using the metrics, these are the test results :
Denoising
Non local mean
PIQE Scores : 61.60
Works by averaging pixel values in an image to reduce noise based on their similarityThis allows it to preserve details in the image while effectively reducing noise
Boxblur
PIQE Scores : 75.27
Averaging the value of the pixels in a square box extending radius pixels in each direction
Average
PIQE Scores : 76.45
Convolving an image with a normalized box filter which takes the average of all the pixels under the kernel area and replaces the central element
Gaussian
PIQE Scores : 74.90
Gaussian blurring utilizes a Gaussian kernel, defined by its width, height, and standard deviations, to reduce Gaussian noise in images
Median
PIQE Scores : 74.32
Takes the median of all the pixels under the kernel area and the central element is replaced with this median value. Highly effective against salt-and-pepper noise in an image
Bilateral
PIQE Scores : 75.88
Reduces noise while preserving edges by utilizing Gaussian filters in both spatial and pixel intensity difference domains considering only pixels with similar intensities for blurring
The Non local mean bluring method perform the best for denoising among all the algorithms
Brightness
Image enhance library from pillow
PIQE Scores : 72.66
Original image is blended with a black image of the same size
Fundamental Point operation
PIQE Scores : 76.47
The equation g(i, j) = α ⋅ f(i, j) + β represents a linear transformation applied to each pixel's intensity value The output pixel value g(i, j) is calculated independently for each pixel in the input image f(i, j), based solely on the input pixel's value and two constants, α and β
The Image Enhance Library from Pillow perform the best for Increasing brightness among all the algorithms
Saturation
OpenCV Librabry
PIQE Scores : 80.85
Increasing the saturation of all pixels by 1.5 times.
Image enhance library from pillow
PIQE Scores : 72.94
The original image is blended with its grayscale version and calculates their gray value of each pixel. The gray value is determined by the mean of the original image' grayscale version.
The Image Enhance Library from Pillow perform the best for enhancing saturation among all the algorithms
Action Taken After the Test
After the test, we understand that which of the following algorithms are the best for our product. They will be integrated to our image processing streamline inside the server of the product.
