Image Segmentation & 3D Stereo Vision in MATLAB

SC4061 Computer Vision, Nanyang Technological University Choo Yi Ken | U2240710B | College of Computing and Data Science (CCDS)

A MATLAB lab implementing and comparing three document image thresholding algorithms (Otsu's, Niblack's, and Sauvola's), followed by a from-scratch Sum of Squared Differences (SSD) disparity map algorithm for 3D stereo vision.

---

Overview

Section	Topic	Input Files
2.1.a	Otsu's Global Thresholding	document01–04.bmp + GT
2.1.b	Niblack's Local Thresholding	document01–04.bmp + GT
2.1.c	Sauvola's Normalised Thresholding	document01–04.bmp + GT
3.1.a	Disparity Map Algorithm (SSD)	— (function definition)
3.1.b–c	Stereo Disparity: Corridor	corridorl.jpg, corridorr.jpg, corridor_disp.jpg
3.1.d	Stereo Disparity: Triclopsi	triclopsi2l.jpg, triclopsi2r.jpg, triclopsid.jpg

---

Techniques Implemented

2.1 Image Segmentation

The goal is to segment dark foreground text from bright backgrounds in four degraded document images, evaluated against ground truth binary images. Accuracy is measured by pixel-wise XOR comparison:

accuracy = (1 - error_pixels / total_pixels) * 100

Polarity inversion is automatically handled — the segmentation with fewer misclassified pixels (original vs inverted) is selected.

---

2.1.a — Otsu's Global Thresholding

Finds a single optimal threshold T that maximises the inter-class variance between foreground and background:

T = argmax_T σ²_between(T)

Uses MATLAB's graythresh (Otsu) and imbinarize. Visualisation includes a pixel intensity histogram with the threshold marked in red, and a 2×2 subplot comparing the original, segmented, ground truth, and difference images.

---

2.1.b — Niblack's Local Thresholding

Computes a spatially-varying threshold at every pixel using the local neighbourhood statistics:

threshold(x,y) = mean(x,y) + k * std(x,y)

Local mean and standard deviation are computed efficiently using imboxfilt. A full grid search is performed over:

• k values: −1.5, −1, −0.5, 0, 0.5, 1, 1.5
• Window sizes: 11, 51, 101, 301, 501

Results are visualised as a 2D heatmap (k vs window size, coloured by error count), and the best parameter combination's threshold map is displayed as a 3D surface.

---

2.1.c — Sauvola's Normalised Thresholding

An improvement on Niblack that normalises the standard deviation term by a dynamic range constant R, reducing over-segmentation in low-contrast regions:

threshold(x,y) = mean(x,y) * (1 + k * (std(x,y) / R - 1))

A three-parameter grid search is performed over:

• k values: −1.5, −1, −0.5, 0, 0.5, 1, 1.5
• R values: 64, 96, 128, 160, 192
• Window sizes: 11, 51, 101, 301, 501

The full parameter space is visualised as a 3D scatter plot coloured by error count, with the best-performing combination marked as a red star.

---

3.1 3D Stereo Vision

3.1.a — Disparity Map Algorithm (SSD)

A custom block-matching disparity algorithm is implemented from scratch. For each pixel in the left image, an 11×11 template is matched against candidate positions in the right image within a maximum disparity window of 15 pixels (searching left only, as expected for rectified stereo pairs). The best match is found by minimising the Sum of Squared Differences (SSD):

SSD = Σ(I_right - G_left)² = ΣI² + ΣG² - 2·Σ(I·G)

Squared image buffers (img²) are precomputed for efficiency.

3.1.b–c — Corridor Stereo Pair

The left/right corridor images (corridorl.jpg, corridorr.jpg) are loaded, converted to grayscale, and passed into the disparity algorithm with an 11×11 template. The resulting disparity map is displayed using a gray colormap with range [−15, 15] and compared visually against the ground truth (corridor_disp.jpg).

3.1.d — Triclopsi Stereo Pair

The same pipeline is rerun on a second stereo pair (triclopsi2l.jpg, triclopsi2r.jpg) and compared against its ground truth (triclopsid.jpg), demonstrating the algorithm's generalisability across different scenes.

---

Requirements

• MATLAB with the Image Processing Toolbox

---

Setup & Usage

1. Place all input images in the same directory as sc4061lab2.m:

document01.bmp   document01-GT.tiff
document02.bmp   document02-GT.tiff
document03.bmp   document03-GT.tiff
document04.bmp   document04-GT.tiff
corridorl.jpg    corridorr.jpg    corridor_disp.jpg
triclopsi2l.jpg  triclopsi2r.jpg  triclopsid.jpg

2. Open sc4061lab2.m in MATLAB and run it section by section using Run Section, or run the entire script at once.

Note: The Niblack and Sauvola sections perform exhaustive parameter grid searches and may take several minutes to complete depending on image size and hardware.

Note: The disparity map computation (Section 3.1) is computationally intensive due to the pixel-level nested loop. Runtime scales with image resolution and maxDisparity.

---

Author

Choo Yi Ken | U2240710B
College of Computing and Data Science (CCDS)
Nanyang Technological University