Computer Vision · Beginner interview

Computer vision is the field of enabling computers to interpret and understand visual information — images and video. Applications include image classification, object detection, facial recognition, medical imaging, autonomous driving, OCR, and quality inspection. The goal is to extract meaning (what's in the image, where, and what's happening) from raw pixels.

An image is a grid of pixels, each with intensity values. A grayscale image is a 2D array of single values (0–255); a color image is a 3D array with channels — typically Red, Green, Blue — so its shape is height × width × 3. Models consume these numeric arrays, which is why preprocessing like resizing and normalization matters.

Common steps: resizing to a fixed input size, normalizing pixel values (scaling 0–255 to 0–1 or standardizing with dataset mean/std) so training is stable, and sometimes converting color spaces or denoising. For training you also augment. Consistent preprocessing between training and inference is essential to avoid skew.

Filtering applies a small kernel over the image to produce effects — blurring (averaging/Gaussian) smooths noise, sharpening enhances detail. Edge detection (Sobel, Canny) finds where intensity changes sharply, outlining object boundaries. These classical operations are convolutions with fixed kernels, and they're the conceptual basis for what CNNs learn automatically.

Convolution slides a small matrix (kernel/filter) across the image, and at each position computes a weighted sum of the overlapping pixels, producing a new value. Different kernels detect different patterns — edges, corners, textures. In CNNs the kernels are learned rather than hand-designed, so the network discovers which patterns matter for the task.

Images can be represented in different color spaces — RGB is standard, but HSV (hue, saturation, value) separates color from brightness, which is useful for color-based segmentation or lighting robustness. Converting to grayscale reduces three channels to one, cutting computation when color isn't informative (e.g. some OCR or edge tasks). Choosing the representation suited to the task can simplify the problem.

A Convolutional Neural Network learns hierarchical visual features using convolutional layers that apply learnable filters across the image. It suits images because it exploits spatial structure: weight sharing makes it parameter-efficient and translation-invariant (a feature is detected anywhere), unlike a dense network that would need enormous parameters and ignore locality. Stacked convolutions build from edges to shapes to objects.

Convolution layers apply learned filters to detect local patterns, producing feature maps. Pooling layers (e.g. max pooling) downsample those maps, reducing spatial size and computation while adding robustness to small translations. Alternating them progressively shrinks spatial dimensions and grows the number of feature channels, capturing increasingly abstract features toward the final classification.

A feature map is the output of applying a filter across the input — it highlights where that filter's pattern appears. Early layers produce maps for low-level features (edges, colors, textures); deeper layers combine these into higher-level features (shapes, object parts, whole objects). This hierarchy, learned automatically, is what lets CNNs recognize complex objects from raw pixels.

Image classification assigns a single label to the whole image ('this is a cat'). Object detection locates objects with bounding boxes and labels ('cat here, dog there'). Segmentation labels every pixel — semantic segmentation by class, instance segmentation by individual object. They increase in spatial granularity and difficulty, and you pick based on whether you need the category, the location, or precise shape.

Augmentation generates new training images via label-preserving transformations — flips, rotations, crops, scaling, brightness/contrast/color jitter, and more. It enlarges and diversifies the dataset, making the model robust to variations it will see in the real world and reducing overfitting, which is especially valuable when labeled images are scarce. You apply it during training only.

Transfer learning reuses a model pretrained on a large dataset like ImageNet, whose early layers have learned general visual features, and adapts it to your task — typically replacing the final layer and fine-tuning. It dramatically reduces the data and compute needed and usually outperforms training from scratch, which is why it's the default approach for most practical vision problems.