Coursework from 1MD110 Introduction to Image Analysis (period 1, year 1) of the M.Sc. in Image Analysis and Machine Learning at Uppsala University.
The course covered classical image analysis: point operations and histograms, thresholding and segmentation, spatial filtering, multispectral imagery, and image registration. Everything here is MATLAB, using the Image Processing and Computer Vision toolboxes.
These are the loose scripts at the repo root. Most of them run on the toy images in images/.
assignment1.m — histogram flattening done by hand. It walks every grey level, and where a level holds more pixels than the target uniform count it pushes the surplus up to the next level, redistributing intensities toward a flatter histogram instead of calling a built-in.
assignment3.m — color segmentation. Splits an RGB image into its channels and into HSV, runs Otsu thresholding (graythresh / imbinarize) on each channel, and shows the six binary masks side by side. It then hands the image to the interactive segmentation tools below.
colorseg_RB.m / colorseg_SV.m — interactive 2D-histogram segmentation. You draw a polygon over the joint Red-Blue (or Saturation-Value) histogram and the pixels falling inside that region get highlighted live as an overlay. Useful for picking out a colour cluster by eye.
assignment4.m — multispectral Landsat data (landsat_data.mat). Builds a false-colour composite from bands 4/1/3 and shows all bands as a montage.
assignment5.m — grayscale conversion, centre-crop to a square, resize to 128×128, then a hand-written mean (box) filter with a sliding window. Subtracting the blurred image from the original gives a simple high-pass / sharpening result.
IterativeTresh.m — iterative global thresholding. Starts from a guess, splits pixels into two groups by the current threshold, and sets the new threshold to the mean of the two group means until it converges. This is k-means with two clusters; it's the clean standalone version.
assignmentv2.m — the longer thresholding script. It runs the same iterative threshold on rice.png, then implements Otsu's method from scratch (sweeping every threshold and maximising the between-class variance) and plots that variance curve. The from-scratch threshold is then checked against MATLAB's built-in graythresh on hand2BW.png, with both binarised results shown side by side. The graythresh value is in [0, 1], so it's scaled back to [0, 255] before the comparison. It reads hand2BW.png from images/ in this repo.
lab4-image-registration/ aligns a moving image to a fixed one with a rigid transform, and scores each method by the mean landmark distance before and after registration (landmarks are in the *.csv files alongside the images in data/). Three approaches:
intensity_based_registration_mono.m— monomodal, regular-step gradient descent with a mean-squares metric over a 5-level pyramid.intensity_based_registration_multi.m— multimodal, (1+1) evolutionary optimizer with Mattes mutual information, for images of differing modality where intensities don't match directly.feature_based_registration.m— detects SIFT keypoints in both images, matches descriptors, and estimates the transform from the matched pairs.
plot_image_and_points.m is the shared helper for overlaying landmark points (including ones that fall outside the image).
lab4-motion-tracking/ holds VideoUtilities.m, the frame-by-frame skeleton for reading a video, annotating tracked positions, and writing the result back out, plus a binary mask (bwmask.png). The course-provided source video isn't redistributed here.
Open the repo in MATLAB with the Image Processing and Computer Vision toolboxes installed. The root scripts are mostly self-contained — cd into the repo so the relative images/ and landsat_data.mat paths resolve. The lab 4 registration functions take the data folder and the fixed/moving image and landmark filenames as arguments, e.g.
T = intensity_based_registration_mono('lab4-image-registration/data', ...
'z1f.png', 'z1m.png', 'z1f.csv', 'z1m.csv');Each registration function prints the mean landmark distance before and after registration, so you can read off how much each method closed the gap, and shows the before/after fused overlays. The data folder pairs a fixed image (*f.png) with a moving one (*m.png) and their landmark CSVs; the b* set is the brain images and the z* set the others, with *lab* variants for the multimodal case.
The assignment solutions are my own: the histogram flattening, the channel/Otsu segmentation, the spatial filter, the iterative and from-scratch Otsu thresholding, and the choices and tuning in the lab 4 registration functions. The registration data set, plot_image_and_points.m, the VideoUtilities.m tracking skeleton, and the lab-4 source video came with the course.
The MIT license covers the code I wrote. The course-provided data and skeleton files stay under their original terms.