Lecture 4b: Image Processing—Spatial Filtering

The lecture explains the application of spatial filtering techniques on digital images, focusing on the foundational mechanics of correlation and convolution operators. Within this framework, smoothing methodologies are analyzed by contrasting linear box blur filters with Gaussian operators, where a normal distribution determines neighborhood weight coefficients to achieve scale-invariant noise attenuation. The mathematical significance of partial derivatives is established as the core mechanism for feature extraction and edge detection, demonstrating how directional intensity gradients derived from first-order derivatives isolate step transitions. This provides the mathematical baseline to construct specialized kernels from derivative principles, culminating in the derivation of the Laplacian filter. As an isotropic operator grounded in second-order partial derivatives, the Laplacian maps localized intensity rate changes to identify zero-crossings, establishing a highly robust digital processing pipeline for omnidirectional boundary detection.