Computer vision, a field of artificial intelligence, involves the development of algorithms and models that allow machines to understand and process images and videos similarly to how humans do. This can include tasks like object detection, image classification, image segmentation, and facial recognition, etc.

Pre-requisite: Basic Machine Learning concepts

Computer Vision Pipeline

Step 1: Image/Video Acquisition

This is the first step in the pipeline, where images or videos are captured using cameras or other sensors. The quality of the images is crucial for the accuracy of the subsequent steps.

You can also choose to use the pre-existing datasets. Some of the most common datasets for computer vision tasks are given in the last section.

Step 2: Preprocessing

In this stage, the raw image/video data is prepared for further analysis. This may involve tasks such as:

• Noise reduction: Removing unwanted noise or artifacts from the images.

• Geometric transformations: Correcting distortions or perspective shifts.

• Normalization: Ensuring that all images have a consistent format and scale.

The preprocessing steps are done so that the images easier to work with or analyze.

Step 3: Feature Extraction

This stage involves identifying and extracting relevant features from the image that can be used for further analysis or classification. Some common features include:

• Edges: Boundaries between regions with different intensities.

• Corners: Points where edges intersect.

• Textures: Patterns or structures within an image.

• Descriptors: Numerical representations of image features

Step 4: Prediction/Recognition Model

Based on the task you are trying to solve you need to make the Machine Learning model accordingly.

For object detection, this involves identifying the location and boundaries of objects within an image. This can be achieved using techniques such as:

• Sliding window: A window is moved across the image, and a classifier is applied at each location to determine if an object is present.

• Region-based methods: Regions of interest are identified and classified.

• Deep learning-based methods: Convolutional neural networks (CNNs) are trained to directly detect objects in images.

For Object Recognition and classification, once objects have been detected, they can be recognized and classified. This involves assigning a label or category to each object. This can be done using:

• Traditional machine learning algorithms: Such as support vector machines (SVMs) or random forests.

• Deep learning models: Such as CNNs or recurrent neural networks (RNNs).

There could be some post processing needed to refine the results and preparing them for further use. This may involve tasks such as:

• Non-maximum suppression: Removing duplicate detections.

• Tracking: Following objects over time in videos.

• Visualization: Displaying the results in a human-understandable format.

Step 5: Model Evaluation

Evaluating a computer vision model is crucial to ensure its performance and reliability in real-world applications.

• Accuracy: The proportion of correct predictions.

• Precision: The proportion of true positive predictions among all actual positive instances.

• F1-score: The harmonic mean of precision and recall.

• Mean Average Precision (mAP): A common metric for object detection tasks.

• Intersection over Union (IoU): Measures the overlap between predicted and ground truth bounding boxes.

• Confusion Matrix: A table that summarizes the performance of a classification model.

• ROC Curve: A graphical plot that illustrates the performance of a binary classifier.

• Precision-Recall Curve: A graphical plot that shows the trade-off between precision and recall.

Step 6: Model Deployment

Once a model has been evaluated and deemed satisfactory, it can be deployed into a production environment.

Model can be deployed on cloud platforms like AWS, GCP, Azure, and other cloud providers which offers a range of services for deploying machine learning models.

Computer Vision Research Areas

Research Areas based on Task

Image Processing

• Image enhancement (noise reduction, sharpening, contrast adjustment)

• Image restoration (deblurring, super-resolution)

• Image segmentation (object/region identification)

• Image registration (aligning images)

Object Detection and Recognition

• Object detection (locating and classifying objects in images)

• Object tracking (following objects in videos)

• Facial recognition (identifying and verifying individuals)

• Action recognition (understanding human or animal actions)

• Gesture recognition (interpreting hand and body movements)

Image and Video Understanding

• Image captioning (generating text descriptions for images)

• Visual question answering (answering questions about images)

• Video summarization (creating concise representations of videos)

• Video action recognition (understanding actions in videos)

Image and Video Generation

• Image generation (creating new images)

• Image editing (modifying existing images)

• Video generation (creating new videos)

• Style transfer (transferring style from one image to another)

Research Areas based on Application

Medical Image Analysis

• Medical image segmentation

• Disease detection (cancer, Alzheimer's, etc.)

• Medical image registration

• Computer-aided diagnosis

Autonomous Vehicles

• Object detection and tracking

• Lane detection

• Pedestrian detection

• Traffic sign recognition

Robotics

• Visual servoing

• Object manipulation

• Simultaneous Localization and Mapping (SLAM)

• Human-robot interaction

Surveillance

• Person re-identification

• Anomaly detection

• Crowd analysis

Augmented Reality

• Object tracking and recognition

• 3D reconstruction

• Image registration

Research Areas based on Techniques

Deep Learning

• Convolutional Neural Networks (CNNs)

• Recurrent Neural Networks (RNNs)

• Generative Adversarial Networks (GANs)

• Transformers

• Self-supervised learning

• Reinforcement learning

Traditional Computer Vision

• Feature extraction (SIFT, SURF, HOG)

• Image matching

• Stereo vision

• Structure from motion

Emerging Research Areas

• Explainable AI

• Adversarial Attacks and Defenses

• Few-shot Learning

• Unsupervised and self-supervised learning

• Video Understanding

• 3D Vision and Augmented Reality

• Large Scale Multi-modal ML Models

Some Famous CV Datasets

After words

Check out more articles in the Computer Vision Series