Foundations of Computer Vision

- Overview

Computer vision empowers machines to interpret visual data from cameras and images by utilizing AI, algorithms, and neural networks to detect pixels, patterns, and objects, often surpassing human speed and accuracy in tasks like industrial inspection. 

While human vision uses context and experience, computer vision relies on high-speed data processing to overcome the "semantic gap," acting as a key component of artificial intelligence (AI).

Ultimately, computer vision bridges the physical world and computer intelligence, allowing machines to "see" and interpret their surroundings.

1. Key Aspects of Computer Vision: 

• Mechanism vs. Biology: Unlike the human eye-brain system, computer vision relies on cameras to capture data, which algorithms then analyze as pixels for color, brightness, and patterns.
• Overcoming the Semantic Gap: The main challenge is translating digital pixel data into high-level semantic meaning, such as identifying a specific object in a scene.
• The Power of Neural Networks: Artificial neural networks represent a significant breakthrough, allowing machines to simulate human-like recognition by analyzing components in a specific sequence.
• Advantages over Human Vision: Computer vision systems can process thousands of products or images per minute, providing unmatched speed, accuracy, and efficiency in tasks like manufacturing quality control.
• Applications: It is widely used for object detection, medical image analysis, surveillance, and autonomous navigation.

2. Differences Between Human and Computer Vision:

• Context and Adaptability: Humans intuitively understand context and can recognize objects in varying lighting or from new angles, whereas computer vision often struggles with variations not present in their training data.
• Speed vs. Understanding: Computers can analyze complex visual inputs in a fraction of a second, but they lack the deep understanding of the physical world that humans possess.
• Training Method: Computer vision requires massive datasets for training (supervised learning), while human vision develops through observation and experience.

Convolutional Neural Networks (CNNs) are specialized deep learning (DL) models that revolutionized computer vision by automatically extracting hierarchical visual features - such as edges, textures, and complex shapes - using brain-inspired layers (convolution, pooling, fully connected). 

By leveraging GPU computing on large datasets like ImageNet, they excel at tasks like image classification, object detection, and medical imaging.

CNNs remain essential for computer vision because they are robust to input variations, efficient in training on GPUs, and highly effective for automated feature engineering, even in the era of visual Transformers.

1. Core Architecture and Function: 

• Convolutional Layers: Use filters (kernels) to identify patterns, creating feature maps that preserve spatial relationships between adjacent pixels.
• Pooling Layers: Downsample the feature maps to reduce computational cost and dimensionality, focusing on key features.
• Fully Connected Layers: Take the high-level features extracted by convolutional layers to perform the final classification.
• Activation Functions (e.g., ReLU): Introduce non-linearity to learn complex patterns.

2. Evolution and Modern Advancements:

• Foundations: Inspired by biology, they evolved from the 1980s Neocognitron to Yann LeCun's LeNet-5 (1989) and exploded with AlexNet (2012).
• Deep Networks (ResNet): Introduced residual connections to solve vanishing gradient problems, allowing for extremely deep networks.
• Efficiency (MobileNet/EfficientNet): Utilize depthwise separable convolutions to reduce parameters and increase computational efficiency.
• Explainability (Grad-CAM): Provides visual explanations to make models more transparent for critical applications like healthcare.

3. Key Applications: 

• Image Classification/Recognition: Identifying objects in images.
• Object Detection: Locating objects within an image.
• Medical Image Analysis: Detecting anomalies in X-rays or MRI scans.
• Facial Recognition: Identifying individuals.

The standard workflow for computer vision follows a three-step process designed to mimic human visual systems and derive meaning from visual data.

These tasks are utilized across industries such as healthcare (medical image analysis), autonomous vehicles (navigation), and manufacturing (visual inspection). 

1. The Three Basic Steps of Computer Vision:

• Acquiring the image/video: Capturing raw visual data from digital sensors such as cameras, scanners, or drones, typically in formats like JPEG or MP4.
• Processing the image: Standardizing, cleaning, and preparing raw data through preprocessing techniques - such as resizing, normalization, and noise reduction - and extracting features to identify patterns like edges, textures, and shapes.
• Understanding the image: Utilizing machine learning and deep learning models to interpret extracted features, resulting in actionable outputs such as labels, bounding boxes, or semantic segmentation.

2. Primary Task Categories for Computer Vision: 

• Image Classification: Assigning a predefined label or category to an entire image (e.g., classifying an image as a "cat" or "car").
• Object Detection: Locating and identifying specific objects within an image or video, often by drawing bounding boxes around them (e.g., detecting pedestrians in traffic footage).
• Image Segmentation: Dividing an image into distinct regions at the pixel level to accurately outline object shapes, often categorized into semantic, instance, or panoptic segmentation.
• Object Tracking: Identifying and following specific objects across multiple frames in a video sequence, crucial for surveillance and autonomous vehicles.
• Face and Person Recognition: Detecting and identifying individuals by analyzing facial features (landmarks) or entire body forms.
• Optical Character Recognition (OCR): Converting text, both printed and handwritten, found in images or scanned documents into machine-readable digital text.
• Scene Understanding/Reconstruction: Interpreting the overall context of a scene and creating 3D models from 2D image data.

Common computer vision problems involve interpreting visual data through different levels of granular analysis: Image Classification assigns a single label to an entire image, Object Detection locates multiple objects using bounding boxes, and Image Segmentation performs precise, pixel-level outlining of objects. 

These tasks are essential for AI applications, ranging from simple tagging to autonomous navigation.

(A) Common Computer Vision Problems: 

1. Image Classification: This task maps an input image to a single label (e.g., "cat" or "dog") representing the main subject. It requires convolutional neural networks (CNNs) or vision transformers to learn features from datasets like ImageNet. 

2. Object Localization and Detection: These tasks identify what objects are in the image and where they are located by drawing bounding boxes around them. Detection often utilizes models such as YOLO or Faster R-CNN to identify multiple objects simultaneously. 

3. Image Segmentation: This is a more precise technique that classifies every individual pixel, outlining the exact boundary of an object.

• Semantic Segmentation: Labels all pixels according to their category (e.g., classifying all "road" pixels).
• Instance Segmentation: Identifies and separates each unique object instance, even within the same class (e.g., distinguishing between different pedestrians).

(B) Key Differences:

• Classification: "What is in this picture?"
• Detection: "What is in this picture, and where?"
• Segmentation: "What is the exact shape of the objects in this picture?".

[More to come ...]