After this lesson you can explain the image classification task and how CNNs approach it; describe the ImageNet benchmark and why it was a turning point; explain what object detection adds to image classification and how YOLO achieves real-time detection; distinguish between classification, detection, and segmentation tasks; and identify concrete maritime applications of each computer vision capability.
Computer vision is the field of AI concerned with enabling machines to derive meaningful information from visual inputs — images, video, and depth sensor data. What changed in 2012 was the combination of large labelled datasets, deep convolutional networks, and GPU computation — turning computer vision into a core enabling technology for robotics, autonomous vehicles, medical diagnosis, and maritime operations.
This lesson builds on the CNN foundations from Lesson 3 and shows how they were applied to progressively harder vision tasks: classification, detection, segmentation, and foundation models.
For maritime professionals: computer vision is the technology enabling autonomous port cranes, hull inspection drones, shipyard weld quality AI, and CCTV-based anomaly detection on vessels and port facilities.
1. The Image Classification Problem
Image classification assigns a label to an entire image from a fixed set of categories. Before deep learning, this used hand-crafted features (SIFT, HOG) fed into SVMs. After AlexNet (2012), CNNs learned features directly from raw pixels, exceeding human performance by 2015. The key enabler was the ImageNet dataset.
2. The ImageNet Moment — How Data Scale Changed Everything
ImageNet (Deng et al., CVPR 2009) — 14M images, 1,000 categories. The ILSVRC ran 2010–2017 and drove a decade of computer vision progress.
Maritime AI insight: scale of labelled data drives model quality. Maritime datasets are orders of magnitude smaller than ImageNet. Transfer learning from ImageNet-pretrained models is the standard approach.
3. Object Detection — Finding Objects in Images
Object detection extends classification: it simultaneously identifies which objects are present and where (bounding boxes). Two-stage detectors (R-CNN, Faster R-CNN) achieved high accuracy but at 5–7 fps — too slow for real-time.
IoU: Measures bounding box overlap vs. ground-truth. IoU > 0.5 = correct detection.
mAP: Mean Average Precision — standard benchmark metric across classes and IoU thresholds.
4. YOLO — Real-Time Object Detection
YOLO (You Only Look Once) (Redmon et al., CVPR 2016) treats detection as a single-pass regression over an S×S grid — predicting bounding boxes and class probabilities simultaneously. Result: 45 fps vs. 7 fps for Faster R-CNN. YOLOv8 (Ultralytics, 2023) is the most widely deployed production model in maritime applications as of 2025.
YOLO 기반 검사 시스템이 선박 건조 프로세스를 어떻게 바꾸고 있는지 실무 관점에서 확인하세요.
5. Beyond Detection — Segmentation
Semantic segmentation assigns a class label to every pixel. Instance segmentation (Mask R-CNN, ICCV 2017) provides separate masks for each individual object instance.
🔍 Semantic Segmentation
Every pixel gets one class label. Used for: waterway detection in autonomous vessel path planning; ice classification; hull coating condition mapping.
🎯 Instance Segmentation
Each object instance gets its own pixel mask. Used for: individual container tracking; crew detection for access control; weld bead segmentation.
6. Foundation Vision Models — SAM and CLIP
7. Maritime Connection — Vision AI and Cyber Risk
During an E26 CRSI assessment, I have seen increasing adoption of AI-based CCTV analytics in port and vessel access control contexts. The typical deployment pattern is: existing CCTV → video stream to shore-based server → YOLO-based detection → alert to security personnel.
The cyber risk question almost universally absent from procurement: "What happens if someone learns which detection model we use and tests adversarial inputs against our camera feeds?" YOLO model weights are public; an attacker who knows which version a port uses can generate adversarial inputs the same afternoon.
The mitigation: treat the model version, training data distribution, and detection thresholds as sensitive configuration data (IACS UR E26 Section 4.1), and include adversarial robustness in the periodic security assessment alongside traditional penetration testing.
- ImageNet (1.2M labelled images, 1,000 classes) enabled the deep learning revolution in computer vision. Transfer learning from ImageNet-pretrained models is standard for maritime tasks.
- Image classification labels a whole image; object detection adds bounding boxes; segmentation assigns class labels to every pixel.
- YOLO achieves real-time detection (45+ fps) via single-pass regression over a grid.
- YOLOv8 is the most widely deployed production detection model in maritime applications as of 2025.
- CLIP enables zero-shot classification; SAM enables prompt-driven segmentation without task-specific retraining.
- Vision AI in maritime is vulnerable to adversarial patches, data poisoning, and camera feed manipulation. Model configuration is sensitive cybersecurity data.
Lesson 5 — From GANs to Generative AI completes PART 1: GAN (Goodfellow 2014) through Diffusion Models to LLMs and multimodal generative systems that define AI in 2025–26.
- Deng et al. (2009). ImageNet. CVPR 2009.
- Redmon et al. (2016). YOLO. CVPR 2016. arXiv:1506.02640.
- Ren et al. (2015). Faster R-CNN. NeurIPS 2015. arXiv:1506.01497.
- He et al. (2017). Mask R-CNN. ICCV 2017. arXiv:1703.06870.
- Radford et al. (2021). CLIP. arXiv:2103.00020.
- Kirillov et al. (2023). SAM. arXiv:2304.02643.
- MITRE ATLAS — atlas.mitre.org
- IACS UR E26 (Rev.3, 2024) — iacs.org.uk
⚓ Join the ShipPaulJobs Community
Join →
Comments
Post a Comment