The process of image processing and converting to black-and-white images, using contrast enhancement techniques and edge detection algorithms
Since the main purpose of contrast enhancement techniques is to improve image quality to support the output of edge detection algorithms, we will start by discussing edge detection algorithms first
Note: The main purpose is for learning, everything is coded from scratch, avoiding the use of built-in functions, which results in longer processing times than usual
An edge detection algorithm refers to a technique used in image analysis and computer vision to identify the locations of significant edges in an image while filtering out false edges caused by noise
In this project, three edge detection algorithms are used: Canny operator, Marr-Hildreth operator (LoG filter), and Sobel operator (The process of each algorithm can be read and understood through the codes and comments in each .py file)
Each method has its own advantages and disadvantages (though in terms of efficiency, the Canny operator is the most robust). The outputs of these three algorithms are then combined to produce the most optimal result
Moreover, since all algorithms are coded from scratch without using built-in functions, the results of the algorithms will differ from those obtained using pre-built functions. However, this difference is not significant and can sometimes even bring certain benefits
- The Canny detection (from scratch) returns a result with more clearly defined facial features
- The Marr-Hildreth results are infact the same, the differences are simply due to a snippet of code in
marr_hildreth_edge_detection.pythat sets points above a certain threshold to 255 and reduces the rest to 0 - The Sobel operator (from scratch) produces a "weaker" result but effectively reduces noise in the area below the face
The final output will be a combination of one or more of the methods mentioned above
Currently, the code uses the majority voting approach, which simply examines all pixels: a pixel is considered an edge pixel in the final result if the majority of methods (1/1, 2/2, 2/3) classify it as an edge pixel
There are still some issues at present:
- Marr-Hildreth and Sobel introduce more noise compared to Canny. As a result, when using majority voting, this noise inadvertently gets included in the final output
- Sobel produces results that resemble a pencil sketch rather than clear edges, which may initially appear to contain more details than Canny. However, the pixel values are distributed across a range from 0 to 255, making it harder to distinguish edge pixels and leading to more noise in the final result than necessary
One potential solution for achieving better results is to incorporate weighting for each method, prioritizing the results from Canny, while Marr-Hildreth and Sobel serve to supplement small details that Canny might miss. However, determining how to assign the weights effectively is a new challenge that requires further testing and evaluation
Another problem is that, with certain images, edge detection algorithms generate too many correct but unnecessary details, the most typical example being hair. This makes Step 2 more difficult and time-consuming
Image contrast enhancement is a pivotal process in digital image processing that aims to improve the visibility and perceptibility of an image by adjusting and amplifying the difference in the brightness and color of the elements within the image
In this scenario, this manipulation ensures that the distinct features within an image are more easily distinguishable, helping the edge detection algorithm return better results
The file contains three contrast enhancement techniques: IAGCWD (Improved Adaptive Gamma Correction with Weight Distribution), CLAHE (Contrast Limited Adaptive Histogram Equalization), and LA (Luminance Adaptation) (You can read and understand each method through the code and comments in the .py files and the references)
Each technique has its own specific utility, however, this discussion will focus solely on the impact of these techniques on the results of edge detection algorithms, in this case, the Canny operator:
- OG Canny (apply canny to original image): Detects major edges clearly but struggles in low-contrast areas, missing finer details
- IAGCWD: Significantly enhances brightness and highlights finer details, especially in regions with low contrast. It outperforms OG Canny in detecting smaller and less prominent edges
- CLAHE: Improves edge detection in complex areas by enhancing local contrast. However, it may introduce slight noise in areas with less information
- LA: Balances brightness effectively, ensuring edges are uniformly visible across the image. While it maintains overall structure well, it lacks the sharpness provided by IAGCWD or CLAHE
Overall conclusion: IAGCWD is effective for enhancing details and detecting subtle edges, while CLAHE is excellent for local contrast improvement but slightly prone to noise, and LA is well-balanced but less sharp in edge detection
Below are the main articles and documents on contrast enhancement techniques, which have significantly contributed to the development of the code segments
- Gamma Correction (IAGCWD):
- Histogram Equalization (CLAHE):
- Luminance Adaption: