Exam-Type Questions Flashcards
Outline some of the key reasons why we would use Machine Vision systems for industrial applications such as automated inspection or process control.
- High accuracy & precision: Machine vision systems use advanced algorithms and sensors to capture and analyse images with high accuracy and precision. This allows them to identify small defects or variations in objects with high accuracy, making them ideal for applications such as automated inspection or quality control.
- High speed & throughput: Machine vision systems can process images and make decisions quickly, allowing them to handle high volumes of objects or samples in a short amount of time. Suitable for automated sorting or material handling.
- Robustness & reliability: Machine vision systems are designed to be robust and reliable, with built-in-error-correction and fault-tolerance mechanisms. This allows them to operate continuously without the need for frequent maintenance or downtime, making them ideal for applications that require 24/7 operation.
- Flexibility & adaptability: Machine vision systems can be easily configured and programmed to perform a wide range of tasks, making them highly flexible and adaptable to different applications and environments. This allows them to be used in a variety of industrial settings, from manufacturing to logistics to agriculture.
What are the key differences between Human and Machine vision?
- Speed & accuracy: Machine vision systems can process images and make decisions much faster and more accurately than humans. This allows them to handle high volumes of objects or samples in a short amount of time, with high accuracy and precision.
- Robustness & reliability: Machine vision systems are designed to be robust and reliable, with built-in-error-correction and fault-tolerance mechanisms. This allows them to operate continuously without the need for frequent maintenance or downtime, making them ideal for applications that require 24/7 operation.
- Consistency and repeatability: Machine vision systems are capable of performing tasks consistently and repeatably, without the influence of factors such as fatigue, boredom, or bias. This allows them to produce consistent and reliable results, regardless of the environment or application.
- Flexibility and adaptability: Machine vision systems can be easily configured and programmed to perform a wide range of tasks, making them highly flexible and adaptable to different applications and environments. This allows them to be used in a variety of industrial settings, from manufacturing to logistics to agriculture.
What are the key issues with the Canny edge detector?
- sensitivity to noise: Small variations in the intensity values of the pixels can affect the results of the gradient calculation and thresholding. This can cause false edges to be detected or real edges to be missed, reducing the accuracy of the edge detection.
- Lack of scale-invariance: Meaning that it may not work well for objects at different scales or distances from the camera. This can limit the applicability to a wide range of images and applications.
- Manual tuning of parameters: Requires manual tuning of the standard deviation, upper threshold, and lower threshold values to achieve the best results. This can be time consuming and error prone, and it may not be possible to find a single set of values that works well for all images.
- Computational complexity: Involves multiple steps such as noise reduction, gradient calculation, non-maximum suppression, and hysteresis thresholding. This can make the algorithm slow to run on large or high-resolution images, and it may not be suitable for real-time applications or systems with limited computational resources.
- Limitations of the Sobel operator: Based on a finite difference approximation of the derivative. This approximation can introduce errors and artefacts in the gradient image, which can effect the accuracy and relatability of the edge detection
Clearly indicate the role of the three key control parameters in the canny edge detection (standard deviation, upper and lower threshold)
Key controls are the standard deviation of the gaussian filter, the upper threshold for hysteresis thresholding, and the lower threshold for hysteresis thresholding.
Standard deviation controls the amount of noise reduction in the image. A higher standard deviation will result in more smoothing and less noise.
The upper threshold controls the sensitivity of the edge detection, and a higher upper threshold will result in fewer edges being detected.
The lower threshold controls the connectivity of the edges, and a higher lower threshold will result in fewer disconnected edges being detected.
Disadvantages of Roberts and Prewitt edge detection methods
- Both sensitive to noise, as small variations in the intensity values of the pixels can affect the results of the convolution and thresholding. This can cause false edges to be detected or real edges to be missed, reducing the accuracy of the edge detection.
- Lack of scale invariance, meaning that they may not work well for objects at different scales or distances from the camera. This can limit their applicability to a wide range of images and applications.
- Both methods require manual tuning of the threshold value to determine which values are considered edges, which can be time-consuming and error-prone.
Why is region labelling of a binary image a useful image analysis tool?
it allows us to identify and classify the objects or regions in a binary image. This can be useful for a variety of applications, such as object recognition, segmentation or tracking.
Region Labelling is generally done based on one of two possible implementations, what are these two approaches?
By area
By location
Which labelling solution is considered better - 4-connected or 8-connected region labelling by location?
Depends on the specific application and the desired results. In general, 4-connected region labelling is considered to be more restrictive, as it only considers pixels that are directly adjacent to each other. Can be useful for applications where it is important to identify distinct objects that are separated by gaps / boundaries, as it will not merge these objects into a single region.
8-connected region labelling is considered to be more flexible and permissive, as it considers pixels that are diagonal to each other as well as pixels that are directly adjacent. This can be useful for applications where it is important to identify regions that are connected or continuous, as it will merge these regions into a single region.
Limiting factors of the 3x3 Chamfer Algorithm based Distance Transform
Size & resolution: suitable for small or low-resolution images. May not be accurate or efficient for larger or high-resolution images, as it may not capture the fine details or structures in the image, and it may require a significant amount of computational resources. This can limit the applicability and performance of the algorithm.
Accuracy & efficiency: May not be the most accurate or efficient for all images or application. e.g. may not be able to accurately capture the distances in images with complex or irregular shapes, or it may be computationally expensive for large or high-resolution images. Affects the reliability and performance of the algorithm.
In order to save memory, you decide to use only one image, A, in a 3x3 mask based filtering process (i.e. A is the input/source and output/destination image). Explain why this will not work. For what sort of operations could you use this single image approach?
Using only 1 image, A, in a 3x3 mask based filtering process will not work because the filter requires at least 2 images: the input/source image, which provides the pixels that are used in the calculation, and the output/destination image, which stores the result of the calculation.
There are some operations where it is possible to use a single image as both the input/source and output/destination. e.g. Some types of image processing operations, such as thresholding and colour space conversion, only require a single input image and do not involve any complex calculations that would require an additional image.
Explain, in detail, how the Double (Hysteresis) Thresholding based approach to Reconstruction by Dilation can be used to extract the watermark from the grey scale image illustrated in Figure Q2.3. You can assume that the image has been corrected for any illumination variations. Clearly explain how each threshold value is chosen with a view to optimizing the reconstruction process.
Works by applying a double thresholding operation to the grey scale image, where the first threshold separates the pixels of the watermark from the background, and the second threshold separates the pixels of the background from the watermark.
- Compute the gradient image of the greyscale image. Done using a gradient operator, such as the Sobel operator, which computes the gradient magnitude and direction of each pixel in the image.
- Apply a double threshold to the gradient image. The 1st threshold separates the pixels of the watermark from the background, and the 2nd separates the pixels of the background from the watermark.
- Perform a reconstruction by dilation on the gradient image. This will fill in the gaps and holes in the watermark, and smooth the boundaries, resulting in a more complete and continuous representation of the watermark.
- Threshold the reconstructed gradient image to create a binary watermark image. Pixels with values above the threshold value are considered to be part of the watermark, and are set to 1. Pixels with values below the threshold value are considered to be part of the background and are set to 0.
The choice of threshold values in the double thresholding operation is critical to the success of the reconstruction by dilation process. In general, the 1st threshold should be chosen to be high enough to separate the watermark and background pixels, but not so high that it excludes too many pixels of the watermark.
Compare and contrast Histogram Stretching and Histogram Equalization approaches to image processing.
2 techniques for enhancing the contrast and the visibility of an image by modifying its intensity distribution. Both techniques operate on the histogram of the image, which is a graph that shows the number of pixels at each intensity level.
Both are similar in that they both operate on the histogram of the image and aim to enhance its contrast and visibility. They differ in their approach and assumptions, as histogram stretching assumes a linear relationship between the original and the transformed intensities, while histogram equalisation assumes a uniform distribution of the transformed intensities. This makes histogram stretching more suitable for images with a limited range of intensities, while histogram equalisation is more suitable for images with a broad range of intensities and a non-uniform distribution
Detail the main elements you would expect to find in an industrial machine vision system (i.e. such as a conveyor based machine vision inspection system).
- Image acquisition device: captures the images of the objects on the conveyor belt and converts them into digital signals. Can be a camera, a scanner, can be monochrome, greyscale, or colour, depending on the requirements of the application.
- Image processing and analysis software: processes the digital images, extracts the relevant features, and performs the required analysis, in order to identify and classify the objects on the conveyor belt. Can be based on algorithms, such as filtering, segmentation, recognition, or classification, and can be implemented using programming languages.
- Object handling and control device: receives the output of the image processing and analysis software, and performs the required actions, in order to handle and control the objects on the conveyor belt. Can be a robot, a gripper, a reject mechanism, and can be programmed to perform the desired actions, such as sorting, counting, measuring or rejecting.
- User interface and communication device: provides the interface between the system and the user, and enables the user to monitor, control and communicate with the system. Can be a display, a keyboard, a mouse, and can be connected to the system through wires or wireless communication channels.
- Lighting and optics device: illuminates the objects on the conveyor belt, and captures the images of the objects, with the required brightness, contrast, and resolution. Can be a light source, a lens, a filter, and can be adjusted to optimise the performance of the system, depending on the characteristics of the objects, the camera, and the environment.
What is the key issue with the Marr-Hildreth edge detector? Indicate one possible solution to this issue.
Strong edges required
standard deviation has to be increased (strong smoothing)
less evident edges are suppressed
may cause a significant loss of useful information.
Solution
use multiple scales and aggregate the information between them.
- Computationally intensive since the convolution masks become larger when
increases.
