Computer Vision Lab

Convolution is the core of CNNs. A small matrix called a kernel (or filter) slides across the image. At each position, element-wise multiplication and summation produce a single output value. Different kernels detect different features — edges, textures, shapes. A CNN learns these kernels during training. Try the presets below to see what each kernel "looks for."

Color spaces are different ways to represent color. RGB stores Red, Green, Blue channels — great for displays. HSV separates Hue (color), Saturation (purity), Value (brightness) — used in color filtering. LAB separates Lightness from color — used for perceptual color comparison. Understanding color spaces is essential for image preprocessing in ML pipelines.

Edge detection finds boundaries where pixel intensity changes sharply — a fundamental step in computer vision. Canny is the most popular — it uses gradient magnitude + non-maximum suppression. Sobel detects directional gradients (horizontal or vertical edges). Harris corners find intersection points of edges. These features are used for object detection, tracking, and SLAM.

Histograms show how pixel values are distributed across 0–255. A dark image has values clustered left; a bright image, clustered right. Histogram equalization spreads values evenly — a key preprocessing step in ML to improve contrast and normalize inputs. This directly affects model performance — many CV pipelines include histogram normalization.

Annotation is how you teach AI to see. Object detection models like YOLO, Faster R-CNN, and SSD need labeled data: bounding boxes drawn around objects with class labels. This tool lets you draw boxes, assign classes, and export in YOLO (class cx cy w h normalized), COCO (JSON), or Pascal VOC (XML) format — the 3 industry standards.

Data augmentation prevents overfitting by creating variations of training images. Instead of collecting thousands more images, you apply transformations: flips, rotations, noise, brightness changes, crops, and color jitter. Modern frameworks like Albumentations, torchvision.transforms, and Keras ImageDataGenerator apply these in real-time during training. A model trained with augmentation generalizes much better to unseen data.

Thresholding converts a grayscale image to binary (black & white) by choosing a cutoff value. Otsu's method automatically finds the optimal threshold. Adaptive thresholding uses local neighborhoods — ideal for uneven lighting. Thresholding is the foundation of image segmentation — dividing an image into meaningful regions. Contour detection then finds the boundaries of these regions.

Morphology operations process binary images using a structuring element (kernel). Erosion shrinks white regions (removes noise). Dilation expands them (fills gaps). Opening = erosion → dilation (removes small objects). Closing = dilation → erosion (fills holes). These are critical in post-processing of segmentation masks and preparing images for OCR or object detection.

The Fourier Transform breaks down an image into its sine and cosine components. The output represents the image in the Frequency Domain, where each point represents a particular frequency contained in the spatial domain image. Low-pass filtering removes high frequencies (edges, noise) causing blur. High-pass filtering removes low frequencies, enhancing edges.

Feature Detection is about finding distinct points in an image (like corners or specific textures) that can be reliably found again even if the image is rotated, scaled, or illuminated differently. ORB (Oriented FAST and rotated BRIEF) is a fast, efficient, and open-source alternative to SIFT and SURF. These keypoints are used for panorama stitching, object tracking, and 3D reconstruction.

K-Means Clustering is an unsupervised machine learning algorithm. In image processing, it's used for Color Quantization — reducing the number of distinct colors in an image while preserving its visual appearance. The algorithm groups pixels into 'K' clusters based on their color similarity, and replaces all pixels in a cluster with the cluster's center color.