Q18Digital Image Processing

Spatial domain techniques directly manipulate the pixels of an image, expressed generally as $g(x,y)=T[f(x,y)]$, where $T$ operates on the original image $f$ either at a single pixel (point processing) or over a neighborhood (spatial filtering). They are broadly categorized into point processing (intensity transformations) and spatial filtering (neighborhood processing).

Point Processing

Point processing includes contrast stretching, image negatives ($s=(L-1)-r$), log transformations ($s=c\log (1+r)$, compressing dynamic range for display of Fourier spectra), power-law/gamma transformations ($s=c{r}^{\gamma }$, used for monitor gamma correction), and histogram equalization, which redistributes intensities using the cumulative distribution function to spread out the most frequent values and improve global contrast, particularly in images with poor illumination.

Spatial Filtering

Spatial filtering uses masks (small matrices, typically $3\times 3$) convolved over the image, where the output pixel is a function of the input pixel and its neighbors. Smoothing (low-pass) filters include the mean filter, which replaces each pixel with the average of its neighborhood to reduce Gaussian noise at the cost of blurring edges, and the median filter, an order-statistic (non-linear) filter that replaces each pixel with the median of its neighborhood, which is particularly effective for removing salt-and-pepper noise while preserving edges. Sharpening (high-pass) filters, such as the Laplacian (${\nabla }^{2}f=f(x+1,y)+f(x-1,y)+f(x,y+1)+f(x,y-1)-4f(x,y)$) and gradient-based operators (Sobel, Prewitt), highlight fine details and edges by responding to rapid intensity changes; the Laplacian result is typically added back to the original image to produce a sharpened output, $g(x,y)=f(x,y)+c[{\nabla }^{2}f(x,y)]$.

The selection of the technique depends entirely on the specific application and the type of defect in the image: low-contrast images call for contrast stretching or histogram equalization, noisy images call for smoothing filters matched to the noise type, and images requiring feature emphasis (edges, fine detail) call for sharpening filters, often combined in a sequence — e.g., smoothing to suppress noise followed by sharpening to restore detail — for optimal visual quality.

Worked Example: Mean Filtering

Consider a $3\times 3$ neighborhood of pixel intensities: $\left[\begin{array}{ccc}10& 12& 250\\ 11& 13& 14\\ 9& 10& 12\end{array}\right]$, where the value 250 represents an impulse noise spike. A $3\times 3$ mean filter with uniform weights $1/9$ computes the new center value as the sum of all nine pixels divided by 9: $(10+12+250+11+13+14+9+10+12)/9=341/9\approx 37.9$, which is heavily skewed by the single outlier and still visibly brighter than its true neighborhood value of about 13. A median filter on the same neighborhood sorts the values as $9,10,10,11,12,12,13,14,250$ and picks the middle (fifth) value, 12, completely rejecting the noise spike and recovering a value consistent with the true local intensity. This concretely demonstrates why order-statistic filters outperform linear averaging filters for impulse-type degradations, while linear filters remain preferable for Gaussian-type noise where every pixel carries a small error rather than a few pixels carrying a large one. This distinction is a common source of confusion for beginners, who often assume that a stronger, larger averaging mask always improves image quality, when in fact matching the filter type to the underlying noise model is the deciding factor in producing an acceptable result.