Binarization

A single, fixed intensity value is applied to every pixel in the image. If a pixel's intensity is above the threshold, it is set to white (foreground); otherwise, it is set to black (background). This method is computationally fast and ideal for scenes with uniform, controlled illumination. The optimal threshold is often determined by analyzing the image's histogram to find the valley between the background and foreground peaks.

The threshold value is calculated dynamically for each pixel based on the intensity statistics of its local neighborhood. This is essential for images with non-uniform lighting or varying backgrounds. Common methods include:

• Mean: Threshold is the mean of the local block.
• Gaussian: Threshold is a weighted sum of the local block, giving more importance to central pixels. This prevents over-segmentation in shadowed regions and under-segmentation in brightly lit areas.

An automatic, unsupervised algorithm that determines the optimal global threshold by minimizing the intra-class intensity variance between the foreground and background pixels. It exhaustively searches for the threshold that creates a bimodal histogram with the tightest clusters. This is the standard first-pass approach when the defect-to-background contrast is high and the lighting is stable, removing the need for manual threshold tuning.

Post-processing operations applied to the binary image to clean up noise and refine defect shapes before measurement. Key operations include:

• Erosion: Removes pixels at object boundaries, useful for eliminating small, spurious white noise.
• Dilation: Adds pixels to object boundaries, useful for filling small holes within a detected defect.
• Opening: Erosion followed by dilation, which smooths contours and breaks narrow isthmuses.
• Closing: Dilation followed by erosion, which fuses narrow breaks and fills gaps.

The binary image serves as the direct input for Connected Component Analysis and Blob Analysis. Once the image is binarized, the system identifies contiguous groups of white pixels (blobs) and extracts quantitative features such as area, perimeter, centroid, bounding box, and circularity. The accuracy of these downstream measurements is entirely dependent on the quality of the binarization step; a poor threshold directly corrupts defect size and shape metrics.

A variation where pixels above the threshold are set to black and pixels below are set to white. This is used when the objects of interest (defects) are darker than the background. For example, a black carbon inclusion on a bright metal surface would require inverse binary thresholding to represent the defect as the white foreground for standard blob analysis tools.

An automatic clustering-based image thresholding algorithm that assumes the image contains two classes of pixels (foreground and background). It calculates the optimum threshold separating these two classes so that their intra-class variance is minimal. This is a non-parametric, unsupervised method, meaning it requires no prior knowledge of the image. It is highly effective for images with bimodal histograms but can fail if the defect is extremely small relative to the background or if illumination is non-uniform.

Unlike global methods that apply a single threshold to the entire image, adaptive thresholding calculates a local threshold for each pixel based on a small neighborhood region around it. This is essential for correcting non-uniform illumination, such as vignetting or shadows on a curved metallic surface.

• Mean Thresholding: Uses the mean of the neighborhood area.
• Gaussian Thresholding: Uses a weighted sum of neighborhood values where weights are a Gaussian window. This method prevents a single bright pixel from skewing the local threshold.

A set of non-linear operations applied to binary images to remove noise, isolate individual elements, and join disparate elements. They operate by probing an image with a small shape called a structuring element.

• Erosion: Removes pixels on object boundaries, useful for eliminating small white noise spots.
• Dilation: Adds pixels to object boundaries, useful for filling small holes within a defect.
• Opening: Erosion followed by dilation; removes small objects without significantly changing the area of larger objects.
• Closing: Dilation followed by erosion; fills small holes and gaps without significantly changing the area.

The pixel intensity histogram is the primary tool for determining a binarization threshold. It plots the number of pixels for each tonal value (0-255). A bimodal histogram features two distinct peaks representing the dark background and the bright foreground (or vice versa), with the ideal global threshold located in the valley between them. If the histogram is unimodal or multimodal, global thresholding will fail, necessitating adaptive or multi-level techniques.

An algorithmic application of graph theory that scans a binary image and groups its pixels into components based on pixel connectivity. Once binarization separates the foreground from the background, this operation assigns a unique integer label to each distinct blob (connected group of white pixels). This is the direct bridge between a raw binary image and blob analysis, enabling the extraction of discrete defect measurements like area, perimeter, and centroid coordinates.

A specific local thresholding technique designed for text and document binarization, but highly effective for detecting fine scratches or text on manufactured parts. It calculates the threshold using the local mean and standard deviation: T = m * (1 + k * (s/R - 1)), where m is the local mean, s is the local standard deviation, R is the dynamic range of the standard deviation, and k is a user-defined sensitivity parameter. It excels at preserving fine details against a varying background.