💻 computer science

DeepCork: A Two-Stage Deep Learning Framework for Automated Defect Inspection of Natural Cork Stoppers

DeepCork is a two-stage deep learning framework that combines a ResNet152-based binary classifier with a YOLO-based detector to achieve high-precision, real-time defect inspection of natural cork stoppers while effectively balancing localization accuracy, latency constraints, and predictive uncertainty in industrial manufacturing.

Original authors: Marcos de Souza Oliveira, Jorge Carlos Santos Cardoso, Hugo Dinis Pereirinha da Silva Amaro, Esmaail Albarazi, Armin Gharibi, João Perdiz, Tiago Ventura, Vanessa Fernandes Ferreira, António Silva Ferr

Published 2026-09-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Marcos de Souza Oliveira, Jorge Carlos Santos Cardoso, Hugo Dinis Pereirinha da Silva Amaro, Esmaail Albarazi, Armin Gharibi, João Perdiz, Tiago Ventura, Vanessa Fernandes Ferreira, António Silva Ferreira, André Perrotta, Licinio Roque

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Wine bottles rely on a simple yet sophisticated seal: a stopper made from the bark of the cork oak tree. This material is unique, offering a perfect balance of lightness, resilience, and the ability to keep air out while letting the wine breathe. However, nature is rarely perfect. The bark is a living tissue with its own history, full of natural pores and textures that can easily be mistaken for damage. When manufacturers cut this bark into cylinders to make stoppers, they risk creating hidden cracks or leaving behind signs of insect infestation. These flaws might look like harmless quirks of nature, but they can ruin the seal, allowing oxygen to spoil the wine inside. For centuries, humans have inspected these stoppers by eye, a slow and tiring process where a tired worker might miss a tiny crack or wrongly reject a perfect one. Today, factories need machines that can see as well as a human, but faster and without getting tired, yet the natural irregularity of the cork makes this a difficult task for computers.

Researchers at the University of Coimbra and their partners in the cork industry have developed a new system called DeepCork to solve this problem. Instead of trying to force a single computer program to do everything at once, they built a two-step inspection line that mimics a careful human workflow. The first step acts as a rapid gatekeeper. Imagine a worker quickly scanning a conveyor belt of thousands of stoppers, instantly tossing out the ones that look obviously perfect and setting aside only the ones that might have a flaw. The computer does this using a specialized image-analysis tool trained to recognize the difference between a healthy cork and a damaged one. It looks at the flat top and the cylindrical side of each stopper, deciding with near certainty whether the item is safe or suspicious. If the computer is not absolutely sure a stopper is perfect, it sends it to the second stage.

The second stage is where the detailed work happens. The stoppers that passed the first check but still raised a flag are examined by a more complex system designed to find exactly where the problem is and what kind of problem it is. This system looks for specific types of damage, such as internal cracks that don't break the surface, external cracks that tear the edge, or holes left by insects. It draws a box around the defect, identifying it as a crack or an insect mark, and decides if the stopper should be discarded. This two-part approach is crucial because a single system trying to do both jobs at once often gets confused by the cork's natural texture, mistaking a harmless pore for a serious crack. By separating the tasks, the system first filters out the good ones with high confidence, then uses its full power to analyze only the few that might be bad.

The team tested this system on a massive collection of over 18,000 real images taken from a factory floor. They found that the two-stage method worked remarkably well. For the flat top of the stopper, the system was almost flawless, correctly identifying and locating defects in nearly every case. The cylindrical body, which is harder to inspect because of its curved shape and changing shadows, proved more challenging, but the system still achieved a very high level of accuracy. The researchers also discovered that newer, more complex computer models did not necessarily perform better than slightly older, more established ones. In fact, a specific version of the detection software known as YOLOv8, when paired with the initial filter, provided the best balance of speed and precision. The entire process is fast enough to keep up with a high-speed production line, taking less than eight milliseconds to inspect a single defective stopper, while non-defective ones are cleared even faster.

This work demonstrates that automated quality control for natural materials does not require a single, all-powerful artificial intelligence. Instead, a carefully designed sequence of simpler steps can achieve superior results. The system successfully reduces the risk of false alarms, where a perfect cork is wrongly rejected, and false misses, where a bad cork slips through. While the researchers note that the curved side of the cork remains the most difficult part to inspect perfectly, and that their training data relied heavily on computer-generated variations to fill in gaps, the results are a significant step forward. They have created a transparent and reliable foundation for the future of cork production, ensuring that the natural seal protecting our wine is as trustworthy as the material itself.

Technical Summary: DeepCork

Problem Statement
The manufacturing of natural cork stoppers is inherently prone to defects due to the non-data-driven plank-cutting process and the material's natural heterogeneity. While automated computer vision has replaced manual grading, current deep learning solutions face a critical trade-off: balancing high localization precision with the strict latency constraints (typically <50 ms) required for high-speed industrial production lines. Furthermore, standard "black-box" architectures often suffer from predictive uncertainty, leading to false positives or missed anomalies when distinguishing between natural cork textures (pores, rugosities) and actual structural defects. Existing frameworks frequently lack an explicit division between initial defect presence detection and subsequent defect classification, making them less robust in high-throughput environments.

Methodology: DeepCork Framework
The authors propose DeepCork, a two-stage deep learning framework designed to decouple the inspection process into a rapid binary filtering phase and a precise localization phase. The pipeline processes two distinct geometric views of the cork: the flat top/bottom surfaces ("Topos") and the cylindrical body ("Corpo").

Stage 1: Lightweight Binary Filtering
- Objective: Rapidly separate defective stoppers from non-defective ones with near-absolute certainty to minimize downstream computational load.
- Architecture: The authors benchmarked DenseNet, MobileNetV3, and ResNet152. ResNet152 was selected as the optimal backbone due to its ability to produce a sharp, concentrated probability mass near 1.0, indicating high predictive certainty.
- Logic: A strict confidence threshold ( $\alpha = 0.95$ ) is applied. Only images with $P(\text{Defect}) > 0.95$ proceed to Stage 2. Samples below this threshold are immediately cleared as high-quality. Separate models are trained for Top and Body views to account for geometric variations.
Stage 2: Explainable Defect Localization and Categorization
- Objective: Precisely localize and classify defects for the samples flagged by Stage 1.
- Architecture: The framework utilizes the YOLO (You Only Look Once) family of object detectors. Five versions (v5, v8, v9, v10, v11) were evaluated in "Nano" and "Small" parameter scales.
- Defect Taxonomy: The system classifies four specific defect types:
  - Type A: Internal Cracks (Fenda Interna)
  - Type B: External Cracks (Fenda Externa)
  - Type C: Round Insect Defects (Bicho Redondo)
  - Type D: Head Insect Defects (Bicho Cabeça)
- Logic: A secondary confidence threshold ( $\beta$ ) validates bounding box predictions to suppress false alarms. Independent models are optimized for Top and Body regions.
Data Strategy
- Preprocessing: Circular Hough Transform (CHT) is used to validate the circularity of Top/Bottom views and filter out structurally unreliable samples.
- Balancing: A two-level balancing strategy was employed. First, undersampling was used to balance the binary classification (Defect vs. No Defect). Second, class balancing was applied to the four defect categories in the localization stage to prevent overfitting to dominant classes.
- Augmentation: To address the scarcity of real defect samples (particularly for the "Bicho" categories), the training set was augmented using random rotation, Gaussian noise, Gaussian blur, illumination scaling, and random occlusion, expanding each class to approximately 5,000 instances. The study utilized an initial dataset of 775,673 images, which was partitioned into Training (70%), Validation (15%), and Testing (15%) subsets at the session level. The final testing set comprised 344 images (189 Top and 155 Corpo).

Key Results
The framework was evaluated on the 344-image test set derived from the industrial dataset.

Stage 1 Performance: ResNet152 demonstrated superior robustness, particularly on the challenging Top view, achieving an F1-Score of 0.8764 (Top) and 0.9448 (Body). Its confidence distribution was significantly more concentrated than MobileNetV3 or DenseNet, justifying its selection as the primary filter.
Stage 2 Performance (Localization):
- Top Region: Small-scale YOLO models achieved near-perfect performance. YOLOv8 Small attained an mAP@50 of 0.994 and an F1-Score of 0.988.
- Body Region: This region proved more challenging due to curvature and lighting gradients. YOLOv8 Small achieved the best overall balance with an mAP@50 of 0.912 and an F1-Score of 0.895.
- Spatial Accuracy: YOLOv9 Small recorded the highest overall spatial accuracy with an mAP@50 of 0.956, attributed to its Programmable Gradient Information (PGI) mechanism.
- Comparison: Newer iterations (v10, v11) slightly underperformed compared to v8 and v9, suggesting their speed optimizations may compromise precision on low-contrast natural textures.
Latency and Throughput:
- Stage 1 (ResNet152) operates at ~~4.2 ms per frame (~~238 FPS).
- The combined pipeline for a defective sample (ResNet152 + YOLOv8-Small) results in a worst-case latency of ~7.3 ms (~136 FPS).
- For non-defective samples (the vast majority), the pipeline terminates after Stage 1, running at the maximum speed of the classifier. This dynamic allocation ensures compliance with industrial throughput requirements without compromising inspection performance.

Significance and Claims
The paper claims that DeepCork establishes a reliable and transparent foundation for automated quality assurance in high-throughput industrial inspection. Its primary significance lies in:

Addressing the Accuracy-Latency Trade-off: By decoupling binary filtering from detailed localization, the system avoids the computational cost of running heavy object detectors on every single image, while maintaining high detection rates.
Mitigating Uncertainty: The systematic confidence probability analysis in both stages reduces classification uncertainty and suppresses false positives, a critical requirement for industrial deployment where false alarms can halt production.
Explainability: Unlike single-stage black-box models, the two-stage approach provides clear decision boundaries (via the binary filter) and explicit spatial localization (via YOLO), enhancing the interpretability of the inspection process.

Limitations and Future Work
The authors modestly note two principal limitations:

The Body region remains a performance bottleneck due to perspective distortions and natural lenticels that mimic defects.
The high metrics for the Top region are partly a function of data augmentation, as the original dataset for certain defect categories (specifically "Bicho") was small.

Future work is proposed to expand the real sample base, validate the system on actual industrial edge hardware to confirm the <50 ms latency threshold in real-world conditions, and integrate the pipeline outputs into a VR-based operator training platform to assist human inspectors.

Technical Summary: DeepCork

More like this