22.2. Key Terms and Components

An image can be represented in the form of a matrix where each element represents the intensity (or pixel value) of that part of the image. For grayscale images, it’s a 2D matrix; for RGB images, it's a 3D matrix.

An image matrix is a structured way to represent an image using numerical values. For grayscale images, each pixel's intensity is displayed as a single number in a two-dimensional array (2D matrix) where the rows and columns correspond to pixel positions. In contrast, RGB images, which include color information, are represented as a three-dimensional array (3D matrix) that captures the red, green, and blue intensity values for each pixel. This method allows computers to process and manipulate images mathematically.

Think of an image matrix like a grid of colored tiles. In a grayscale image, each tile represents a different shade of gray, while in an RGB image, each tile consists of a mix of three colors (red, green, blue) in varying intensities. Just like how each square on a grid can be painted a different color, each element in a matrix holds a value that contributes to the overall image.

A smaller matrix (e.g., 3x3 or 5x5) that is used to process the image. It highlights certain features like edges, blurs, or patterns. Example of a 3x3 edge detection filter: [-1, -1, -1] [-1, 8, -1] [-1, -1, -1]

The kernel or filter is a key component in the convolution process. It is a smaller matrix, often much smaller than the image itself, that slides over the image to perform different operations, such as emphasizing certain features. Each element in the kernel is multiplied by the corresponding pixel value it overlaps, and the results are summed to produce a new pixel value. This process can enhance edges, apply blurring effects, or even create special artistic effects, depending on the values in the filter. The example given shows an edge detection filter where the center value is significantly higher than the surrounding values, allowing it to highlight edges.

Imagine using a magnifying glass to examine a patterned fabric. If the magnifying glass is like our kernel, each part of the fabric you focus on will look different depending on how you adjust the glass. Some areas may appear sharper (like edges) while others may blend together. Just as adjusting the magnifying glass can enhance specific details, adjusting the values in a filter can enhance specific features in an image.

The output of applying the convolution operation — a new matrix showing detected features.

The feature map is the result of applying a convolution filter to an image matrix. After the filter has been slid across the image and calculations have been performed at each position, the resulting values are collected into a new matrix called the feature map. This matrix represents the features detected by the filter, making it easier to analyze or process the image further. For example, using an edge detection filter will produce a feature map that highlights the edges found in the original image, simplifying the information for tasks like object recognition.

Think of the feature map like a treasure map. After you explore a landscape (your original image), the treasure map (the feature map) highlights important locations (the features) that are worth noting, such as hills or rivers. Similarly, the feature map condenses all the necessary information from the image into a more manageable form that can be used for further exploration or tasks.

The number of pixels the filter moves each time. A stride of 1 means the filter moves one pixel at a time.

Stride refers to how the filter moves across the image during the convolution operation. If the stride is set to 1, the filter moves one pixel at a time both horizontally and vertically, ensuring that every possible position in the image is covered. Alternatively, if the stride is set to 2, the filter skips every other pixel, which can speed up processing and reduce the output size but may also miss some details. The choice of stride can significantly affect the size of the resulting feature map and the information retained from the original image.

Imagine sliding a piece of paper with a drawing on it across a table. If you move it one inch at a time (stride of 1), you can see every detail; if you move it two inches at a time (stride of 2), you might miss some parts of the drawing. In the same way, adjusting how quickly we slide our filter over the image determines how much detail we capture in the resulting feature map.

Adding extra border pixels (usually zeros) around the image so the filter can fully cover the edges. Helps maintain image size after convolution.

Padding is the technique of adding additional pixels around the border of an image before performing the convolution. This step is important for ensuring that the filter can fully cover the corners and edges of the image, which would otherwise be excluded from processing. Typically, these added pixel values are set to zero (black) in a grayscale image, but they could also take on other values. Padding helps in maintaining the original size of the image in the feature map and prevents information loss at the edges, allowing for accurate feature detection.

Think of padding like wrapping a gift with extra paper to ensure all sides are neatly covered. If you only use the bare minimum, some parts might be left exposed or unwrapped, similar to how edges of an image might not be processed without padding. By adding extra paper (padding), you ensure that every part of the gift (the image) is included in the overall presentation (the convolved output).

--