Computer Vision Evolution and Engineering Paradigm Overview

1.1 Traditional Computer Vision Era: Handcrafted Feature-Driven Paradigm

The history of computer vision (CV) can be traced back to the 1950s, when early researchers first attempted to enable machines to "see" and interpret visual information. In the first 60 years of its development, CV was dominated by handcrafted feature-driven pipelines, which relied on manually designed algorithms to extract key information from images and then use traditional machine learning classifiers for inference. This paradigm is rooted in the core hypothesis that visual information can be represented by a small set of engineered features that are invariant to common variations such as scale, rotation, illumination, and viewpoint.

The milestone of this era is the Scale-Invariant Feature Transform (SIFT) proposed by David Lowe in 1999. SIFT solves the problem of feature invariance by constructing a Gaussian scale space, detecting key points via Difference of Gaussians (DoG), and generating 128-dimensional gradient histogram descriptors for each key point. Its core workflow includes four steps: (1) scale-space extreme point detection to identify scale- and rotation-invariant key points; (2) key point localization to eliminate low-contrast and edge-sensitive points; (3) direction assignment based on local gradient distribution to achieve rotation invariance; and (4) descriptor generation to encode local gradient information. SIFT and its variants (such as SURF, ORB) were widely used in tasks such as image stitching, 3D reconstruction, and object retrieval in the early 21st century.

Another representative work of this era is the Histogram of Oriented Gradients (HOG) proposed by Dalal and Triggs in 2005. HOG counts the gradient direction histogram of local regions of an image, and combined with Support Vector Machines (SVM), it achieves excellent performance in pedestrian detection. In 2001, Viola and Jones proposed the AdaBoost-based face detection framework, which uses Haar-like features and cascade classifiers to achieve real-time face detection, which was widely used in early digital cameras and video surveillance systems.

Despite its historical significance, the handcrafted feature paradigm has inherent limitations that restrict its industrial application. First, feature engineering relies heavily on expert experience, and it is difficult to design universal features that adapt to complex and changeable real scenarios. For example, SIFT has good performance on well-lit, low-blur images, but its recall rate drops by more than 30% in low-light or motion-blur scenarios. Second, the generalization ability of handcrafted features is poor. Even minor changes in the scene (such as the replacement of product models in industrial quality inspection) require redesigning features, which costs a lot of manpower and time. Taking the PASCAL VOC 2007 object detection task as an example, the best mAP (mean Average Precision) of traditional methods is only about 30%, while the deep learning method AlexNet proposed in 2012 directly increased this value to 62%, marking the end of the traditional CV era.

> 📌 Sidebar: Why is SIFT still used in some industrial scenarios? > Although deep learning has become the mainstream, SIFT and ORB are still used in some scenarios with small sample sizes and high requirements for interpretability, such as industrial defect location and visual positioning of robots. Their feature extraction process is interpretable, and no large amount of labeled data is required for training, which is suitable for small-sample scenarios.

1.2 Deep Learning-Driven Computer Vision Era

The outbreak of deep learning in CV originated from the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) 2012. The AlexNet proposed by Hinton's team reduced the top-5 error rate on the ImageNet dataset to 16.4%, which was 10.7 percentage points lower than the 26.1% of the best traditional method before, directly opening the prelude to the deep learning era of CV. Since then, CV technology has developed rapidly, forming a complete technical system around three core tasks: image classification, object detection, and image segmentation.

1.2.1 Evolution of Core Task Technologies

Image Classification

Image classification is the most basic CV task, which aims to assign a global class label to an input image. The evolution of classification models has gone through several stages: from the early AlexNet (5 convolutional layers + 3 fully connected layers) to the 16-layer VGG proposed in 2014, which verified that small 3×3 convolutional kernels stacked in depth can achieve better performance than large convolutional kernels; then to the GoogLeNet proposed in the same year, which introduced the Inception module to improve computational efficiency; and then to the ResNet proposed by He Kaiming in 2015, which solved the problem of gradient disappearance of deep networks through residual connections, making it possible to train hundreds of layers of networks. The 152-layer ResNet reduced the top-1 error rate on ImageNet to 7.3%, which is lower than the human-level performance of about 5% at that time. Later, lightweight classification models such as MobileNet and ShuffleNet were proposed to meet the needs of edge deployment, and Vision Transformer (ViT) based on the Transformer architecture proposed in 2020 further improved the accuracy of classification tasks.

Object Detection

Object detection needs to not only classify objects in the image but also locate their positions via bounding boxes. Its evolution is mainly divided into two branches: two-stage detectors and one-stage detectors. The two-stage detector represented by R-CNN proposed in 2014 first uses Selective Search to generate about 2000 candidate regions, then extracts CNN features for each region, and finally uses SVM for classification and bounding box regression. Its mAP on PASCAL VOC 2007 reached 58.5%, which was 20 percentage points higher than the best traditional method at that time, but the speed was slow, and it took tens of seconds to process one image. The subsequent Fast R-CNN introduced ROI Pooling to share convolutional features of the whole image, and the speed was increased to 1 frame per second, with mAP reaching 70%. Faster R-CNN replaced Selective Search with a Region Proposal Network (RPN) in 2015, realizing end-to-end training, and the speed was further increased to 5 frames per second, with mAP reaching 73.2%.

The one-stage detector represented by YOLO (You Only Look Once) proposed in 2016 treats detection as a regression problem, directly predicts the class and position of the bounding box on the output layer, and achieves a speed of 45 frames per second, with mAP reaching 63.4% on PASCAL VOC 2007. Although its accuracy is slightly lower than that of two-stage detectors, its speed advantage makes it widely used in real-time scenarios. The subsequent YOLOv3, YOLOv4, YOLOv5, YOLOv7, and YOLOv8 have continuously optimized the model structure and training strategy, and the latest YOLOv8x achieves 53.9% mAP on the COCO dataset, with a speed of 150 frames per second on the V100 GPU, which can fully meet the needs of most industrial real-time scenarios.

Image Segmentation

Image segmentation requires pixel-level classification of objects in the image, which is a more fine-grained CV task. It is divided into three sub-tasks: semantic segmentation, instance segmentation, and panoptic segmentation. Semantic segmentation assigns a class label to each pixel, but does not distinguish between different instances of the same class; instance segmentation not only classifies pixels but also distinguishes different instances of the same class; panoptic segmentation combines the two, assigns semantic labels to background pixels, and assigns instance IDs to foreground objects.

The milestone of segmentation tasks is the **Fully Convolutional Network (FC