Paradigm Evolution of Computer Vision and Mathematical Prerequisites

The evolution of computer vision over the past two decades represents one of the most dramatic paradigm shifts in the history of artificial intelligence. For nearly 30 years after the field’s formalization in the 1960s, the dominant research and engineering paradigm relied on handcrafted feature extraction paired with shallow statistical learning models, a framework that achieved notable success in constrained scenarios but fundamentally failed to scale to the complexity of real-world visual perception tasks. The watershed moment arrived in 2012, when the AlexNet convolutional neural network (CNN) achieved a top-5 error rate of 15.3% on the ImageNet Large Scale Visual Recognition Challenge (ILSVRC), a 10.8 percentage point improvement over the previous state-of-the-art result of 26.1% achieved by traditional SIFT + HOG + SVM pipelines. This performance gap was not a marginal improvement but a qualitative leap that exposed the inherent limitations of handcrafted feature engineering and established deep learning as the de facto paradigm for visual perception tasks. This chapter first systematically analyzes the core limitations of traditional computer vision methods, then elaborates the mathematical foundations required to understand and design deep learning-based visual systems, and finally introduces mainstream deep learning frameworks and their core components for computer vision development, laying a solid theoretical and practical foundation for subsequent chapters.

1.1 Core Limitations of Traditional Computer Vision Paradigms

The traditional computer vision pipeline follows a modular, handcrafted design philosophy: first extract discriminative visual features from raw pixels using manually designed algorithms, then train shallow statistical models (e.g., Support Vector Machines (SVM), Random Forests) on these features to perform classification, detection, or segmentation tasks. While this paradigm dominated the field for decades, its inherent architectural limitations prevented it from solving complex real-world visual problems.

1.1.1 Inherent Bottlenecks of Handcrafted Feature Engineering

The core assumption of handcrafted feature design is that human experts can define generalizable feature extractors that capture the essential visual patterns required for downstream tasks. The most widely used handcrafted features in the pre-deep learning era include Scale-Invariant Feature Transform (SIFT), Histogram of Oriented Gradients (HOG), and Speeded Up Robust Features (SURF), each designed to capture specific types of visual statistics.

💡 Side Note: SIFT (Scale-Invariant Feature Transform), proposed by David Lowe in 2004, detects keypoints by comparing Difference of Gaussians (DoG) across multiple scales, achieving invariance to scale, rotation, and affine distortion. While it was the gold standard for feature extraction for nearly a decade, its reliance on handcrafted gradient statistics makes it fundamentally unable to capture high-level semantic patterns, and its computational cost of $O(n^2)$ for n-pixel images makes it impractical for real-time applications on high-resolution imagery.

HOG, proposed by Dalal and Triggs in 2005 for pedestrian detection, computes histograms of gradient orientations in local image cells, achieving robustness to local geometric and photometric transformations. However, HOG’s reliance on local gradient statistics makes it extremely sensitive to object deformation: for example, the HOG descriptor of a standing pedestrian and a bending pedestrian differ by more than 40% in normalized feature space, leading to a 60% drop in detection accuracy for non-upright pedestrians. Similarly, SIFT fails completely on textureless surfaces such as glass, water, and smooth metal, as it relies on gradient magnitude to detect keypoints, leading to zero keypoint extraction for over 30% of natural images in the Microsoft COCO dataset.

Empirical evaluations on standard benchmarks quantify these limitations: on the