The global beer industry, which produced over 1.89 billion hectoliters in 2022, relies heavily on the quality of malt barley to maintain the flavor, aroma, and consistency of its products.

Table of Contents

Malt barley, a specific type of barley optimized for brewing, undergoes a malting process where its starches are converted into fermentable sugars. These sugars are critical for yeast fermentation, the biological process that produces alcohol and carbonation in beer.

Despite its importance, traditional methods of grading barley grains classifying them into categories like Grade 1, Grade 2, Grade 3, or Undergrade—have long been plagued by inefficiency and human error.

A groundbreaking study published in the Journal of Agriculture and Food Research in 2025 introduces an innovative solution: an automated system powered by artificial intelligence (AI) that grades barley grains with remarkable accuracy.

Developed by researchers from Ming Chi University of Technology and Addis Ababa University, this system combines convolutional neural networks (CNNs), a type of deep learning model, with advanced texture analysis.

Using Gabor filters to achieve a testing accuracy of 98.61%, outperforming both human experts and existing machine learning models.

Global Impact of Malt Barley on Modern Brewing Practices

Barley (Hordeum vulgare L.) is the fourth most-produced cereal crop globally, with 146 million metric tons harvested in the 2021/2022 crop year. The European Union leads production, contributing 51.5 million metric tons in 2022/2023.

While food barley is used in bread and animal feed, malt barley is indispensable for brewing. During malting, barley’s starches convert into fermentable sugars through enzymatic reactions, a process critical for yeast to produce alcohol.

The quality of barley directly impacts the efficiency of this process and the final product’s taste. For example, Grade 1 barley, characterized by uniform size, plump kernels, and minimal defects, ensures optimal enzyme activity and sugar yield.

In contrast, Undergrade barley, which may contain shriveled grains, fungal infections, or foreign particles, can lead to off-flavors or incomplete fermentation.

Challenges of Manual Barley Grading in the Brewing Industry

In Ethiopia, a significant barley producer, manual grading remains the norm. However, this method is fraught with challenges, including human bias, time constraints, and high labor costs.

For instance, a Grade 2 grain might be misclassified as Grade 3 due to fatigue or subjective judgment, leading to financial losses for farmers or quality issues for breweries.Traditional grading methods struggle to meet the demands of modern brewing.

Manual inspections, which involve visual scrutiny of grain samples, rarely exceed 85–90% accuracy. Labor costs can consume 20–30% of a brewery’s budget, particularly in regions like Ethiopia, where skilled graders are in high demand.

Deep Learning Solutions for Barley Texture Analysis in Brewing

Earlier attempts to automate grading using classical machine learning models, such as Support Vector Machines (SVMs) or k-Nearest Neighbors (KNN), faced limitations.

These models required manual feature engineering, a process where experts predefined characteristics like color, shape, or size for the system to analyze. For example, an SVM might be trained to recognize Grade 1 barley based on pixel intensity thresholds or geometric measurements.

However, this approach often failed to capture complex textures or subtle defects, such as hairline cracks or slight discolorations, resulting in accuracies of only 70–80%.

The study’s authors recognized that deep learning (DL), a subset of AI that mimics the human brain’s neural networks, could overcome these limitations.

The Role of Texture Analysis in Automated Barley Classification

DL models like CNNs automatically learn patterns from raw data, eliminating the need for manual input. However, while CNNs excel at detecting edges and shapes, they sometimes overlook texture details—a critical factor in barley quality.

To address this gap, the researchers integrated Gabor filters into their CNN model.

Gabor filters, inspired by the human visual system’s ability to detect textures, are mathematical functions used in image processing to highlight specific patterns, such as roughness, smoothness, or directional features.

Innovative Methodology for Automated Barley Quality Assessment

The study began with the collection of 1,200 high-resolution barley images in collaboration with BGI-Ethiopia, a leading brewery. These images were captured under controlled conditions uniform lighting, fixed camera distance, and stable backgrounds to minimize variability.

Each grain was categorized into four classes per Ethiopian Standards Authority (ESA) guidelines: Grade 1 (premium quality, uniform size, no defects), Grade 2 (minor imperfections like slight discoloration), Grade 3 (significant defects such as fungal spots), and Undergrade.

The images were then converted to grayscale, a process that simplifies analysis by reducing color variations, and resized to 224×224 pixels, a standard input size for CNNs.

Next, the Canny edge detection technique, a multi-step algorithm known for its precision, isolated individual grains from the background.

This method involves noise reduction using Gaussian filters, gradient calculation to identify intensity changes, non-maximum suppression to thin edges, and hysteresis thresholding to finalize boundaries.

The result was a set of segmented images with clearly defined grain edges, achieving 95% precision.

Breakthrough Results in AI-Based Barley Classification Accuracy

The preprocessed images underwent texture feature extraction using Gabor filters set at an orientation of 45 degrees (π/4) and a frequency of 0.1 cycles per pixel. These parameters were chosen to highlight diagonal and coarse textures, which are common in barley grains.

The resulting texture maps emphasized patterns like cracks or discolorations, providing the CNN with richer data for classification. For example, a Grade 1 grain might display smooth, uniform textures, while an Undergrade grain could show irregular patches or jagged edges.

The custom CNN architecture featured seven convolutional layers, each designed to learn hierarchical features—from basic edges in early layers to complex textures in deeper layers.

Key components included Leaky ReLU activation, a function that prevents “dead neurons” by allowing small negative values to pass through, and batch normalization, a technique that stabilizes training by standardizing layer inputs.

Dropout layers, set at a rate of 0.4, were also incorporated to reduce overfitting—a phenomenon where the model memorizes training data instead of learning general patterns.

The model was trained using Stochastic Gradient Descent (SGD), an optimization algorithm that adjusts neural network weights to minimize prediction errors.

SGD operates by randomly selecting subsets of data (mini-batches) to compute gradients, making it faster and more memory-efficient than traditional gradient descent.

The learning rate, initially set to 0.001, was dynamically adjusted during training, while a momentum value of 0.9 helped accelerate convergence.

Over 100 training epochs—complete passes through the dataset—the model refined its accuracy, achieving a training accuracy of 99.81% and a testing accuracy of 98.61%.

Only five misclassifications occurred out of 360 tests, as revealed by the confusion matrix, a grid that compares predicted vs. actual classifications.

Transforming Brewery Efficiency with Automated Barley Grading Systems

For context, the study compared its system to three leading CNN models: ResNet-50, DenseNet-121, and VGGNet-19. ResNet-50, a 50-layer residual network, uses skip connections to mitigate vanishing gradients in deep networks.

DenseNet-121 employs dense blocks where each layer connects to every other layer, enhancing feature reuse. VGGNet-19, a 19-layer network, is renowned for its simplicity and depth.

Despite their strengths, these models achieved lower accuracies—73.75%, 70.75%, and 81%, respectively—highlighting the superiority of the proposed texture-CNN hybrid.

The integration of Gabor-filtered textures improved testing accuracy by 1.39% and training accuracy by 2.01% compared to using raw images.

The system’s speed was equally impressive, processing images in milliseconds—a critical advantage for real-time deployment on production lines.

How AI-Driven Barley Grading Enhances Brewery Quality Control

For breweries, this technology offers transformative benefits. Automating grading could save $50, 000-$ 100,000 annually in labor costs, particularly in regions like Ethiopia, where manual inspections dominate.

Consistency is another key advantage. By adhering to ESA standards, the system ensures every batch meets global export requirements. For example, Grade 1 barley guarantees optimal enzyme activity during malting, reducing brewing time by 10–15%.

Farmers also stand to gain, as unbiased assessments prevent undervaluation of high-quality crops. Smallholder farmers, often at the mercy of subjective grading, could receive fair prices for their harvests.

Expanding AI Texture Analysis to Other Agricultural Products

Beyond brewing, the study highlights broader applications for agriculture. Coffee beans, rice, and cocoa—products reliant on texture for quality—could adopt similar systems.

For instance, RiceNet, a CNN-based model, achieved 94% accuracy in classifying rice varieties by analyzing features like grain length and chalkiness.

Similarly, YOLOv8, an object detection model, detected cracks in maize seeds with 99.66% precision using soft X-ray imaging. These advancements underscore AI’s potential to revolutionize food safety and quality control across supply chains.

Future Innovations in AI-Driven Agricultural Quality Control

Despite its promise, the system has limitations. Its reliance on controlled environments uniform lighting and camera stability may hinder performance in real-world settings with variable conditions.

Expanding the dataset beyond 1,200 images, perhaps to 10,000+ samples from diverse regions, could improve robustness.

Future research could explore multispectral imaging, a technique that captures data beyond visible light (e.g., infrared or ultraviolet), to detect defects like mold or insect damage invisible to the naked eye.

Another avenue is developing smartphone apps for field use, enabling farmers to grade crops onsite using mobile cameras. Hybrid models combining CNNs with classical machine learning for morphological analysis measuring grain length or circularity might further enhance accuracy.

Conclusion

In conclusion, this study represents a paradigm shift in agricultural quality control. By merging Gabor filters with deep learning, the researchers have created a system that is not only accurate but also scalable and cost-effective. For breweries, it promises consistent product quality; for farmers, fair pricing; and for consumers, a better pint every time.

As AI continues to permeate agriculture, innovations like this will play a pivotal role in meeting the demands of a growing global population—one grain at a time.

Power Terms

1. Convolutional Neural Network (CNN)
A CNN is a type of artificial intelligence model designed to analyze visual data like images. It uses layers of filters to detect patterns, such as edges or textures, automatically. For example, in the study, CNNs processed barley grain images to classify their quality. CNNs are important because they reduce the need for manual feature selection, making tasks like image classification faster and more accurate.

2. Deep Learning (DL)
Deep learning is a subset of machine learning where algorithms learn from data by mimicking the human brain’s neural networks. It uses multiple layers (hence “deep”) to recognize complex patterns. In this research, DL helped automate barley grading by training the CNN to distinguish between grain qualities. DL is widely used in self-driving cars, voice assistants, and medical diagnostics.

3. Texture Features
Texture features describe the visual “feel” of an image, such as roughness or smoothness. For barley grains, texture indicates quality—smooth grains might be Grade 1, while cracked grains are Under-grade. The study used Gabor filters to extract these features, helping the CNN focus on critical details. Texture analysis is also used in fabric quality checks or satellite image analysis.

4. Gabor Filters
Gabor filters are mathematical tools that detect texture patterns in images by analyzing orientation and frequency. In the study, they extracted barley grain textures to improve grading accuracy. The formula combines a Gaussian (blur) function with a sinusoidal (wave) pattern. For example, Gabor filters are used in fingerprint recognition or tumor detection in medical scans.

5. Malt Barley
Malt barley is a barley variety used primarily in brewing beer. Its quality affects beer flavor, aroma, and fermentation. The Ethiopian Standards Authority (ESA) grades it into four classes (Grade 1 to Under-grade). High-quality malt barley ensures better malting, a process where grains are soaked and dried to produce enzymes for brewing.

6. Brewery Industry
The brewery industry produces alcoholic beverages like beer. It relies on high-quality malt barley to maintain product consistency. Traditional manual grading is slow and error-prone, so automated systems (like the CNN model in the study) help breweries save time and reduce costs. Major producers include the U.S., Germany, and China.

7. Grading System
A grading system categorizes products based on quality. For barley, grades (e.g., Grade 1 or Under-grade) depend on factors like texture, size, and defects. Automated grading using CNNs ensures fairness and speed, replacing human inspectors. Grading is also used for fruits, coffee beans, and diamonds.

8. Segmentation
Segmentation divides an image into parts to isolate objects. In the study, barley grains were separated from the background using edge detection. This step helps the CNN focus on individual grains. Segmentation is used in medical imaging to identify tumors or in self-driving cars to detect pedestrians.

9. Preprocessing
Preprocessing prepares raw data for analysis. For barley images, this included resizing images to 224×224 pixels and converting them to grayscale. Preprocessing ensures consistency and improves model accuracy. Other examples include noise removal in audio files or adjusting brightness in photos.

10. Batch Normalization
Batch normalization is a technique to stabilize and speed up neural network training. It adjusts input data to have a consistent scale. In the study, it reduced training time and improved accuracy. Without it, the model’s accuracy dropped to 89%. It’s like standardizing ingredients in a recipe for consistent results.

11. ResNet-50
ResNet-50 is a pre-trained CNN model with 50 layers. It uses “skip connections” to avoid losing information in deep networks. The study compared its performance to their custom CNN, but ResNet-50 achieved lower accuracy (73.75%). ResNet models are used in image recognition competitions.

12. DenseNet-121
DenseNet-121 is another CNN where each layer connects to all previous layers, enhancing feature reuse. In the study, it scored 70.75% accuracy, less than the proposed model. DenseNet is useful for tasks with limited data, like medical imaging.

13. VGGNet-19
VGGNet-19 is a 19-layer CNN known for simplicity and depth. It achieved 81% accuracy in barley grading but was outperformed by the study’s model. VGGNet is often used as a benchmark in image classification tasks.

14. Dataset Collection
Dataset collection involves gathering and organizing data for training AI models. The study used 1,200 barley images from Ethiopia, split into four grades. High-quality datasets are crucial for reliable AI systems. Examples include facial recognition databases or weather satellite images.

15. Confusion Matrix
A confusion matrix is a table showing how well a model classifies data. In the study, it revealed that 355 out of 360 test images were correctly graded. It helps identify errors, like misclassifying Grade 1 as Grade 2. Doctors use it to evaluate diagnostic tools.

16. Validation Split
Validation split reserves part of the training data to test the model during training. The study used 30% of data for validation to fine-tune the model. It prevents overfitting, where a model memorizes data instead of learning patterns.

17. Transfer Learning
Transfer learning adapts a pre-trained model (like ResNet) to a new task. The study tested transfer learning but found custom CNNs performed better. It’s useful when labeled data is scarce, like using a face-recognition model for pet identification.

18. Edge Detection
Edge detection identifies boundaries in images. The study used the Canny edge detector to outline barley grains. This technique is vital for object recognition in robotics or detecting cracks in materials.

19. Feature Extraction
Feature extraction identifies important patterns in data. The study used Gabor filters to extract texture features from barley grains. Features like color or shape help models make decisions. For example, voice assistants extract pitch and tone from speech.

20. Activation Function (Leaky ReLU)
An activation function decides if a neuron in a neural network should “fire.” Leaky ReLU allows small negative values to pass, preventing dead neurons. In the study, it helped the CNN learn complex patterns. Other functions include Sigmoid (used for probabilities) and Tanh.

21. SoftMax Classifier
SoftMax converts neural network outputs into probabilities. For barley grading, it assigned probabilities to each grade (e.g., 95% Grade 1). It’s used in multi-class tasks like handwriting recognition or spam detection.

22. Stochastic Gradient Descent (SGD)
SGD is an algorithm that adjusts model parameters to minimize errors. The study used SGD with a learning rate of 0.001 to train the CNN. It’s like tuning a radio dial to find the clearest signal.

23. Overfitting
Overfitting occurs when a model performs well on training data but poorly on new data. The study used dropout layers to prevent this. For example, a student memorizing answers instead of understanding concepts would fail new exams.

24. Ethiopian Standards Authority (ESA)
The ESA sets quality standards for Ethiopian agricultural products. It classifies barley into four grades based on size, moisture, and defects. Compliance ensures grains meet brewery requirements, similar to the FDA regulating food safety in the U.S.

25. Accuracy
Accuracy measures how often a model’s predictions are correct. The proposed CNN achieved 98.61% testing accuracy, meaning it graded 986 out of 1,000 grains correctly. High accuracy is critical in fields like healthcare or autonomous driving.

Reference:

Embeyale, D., Chen, Y. T., & Assabie, Y. Automatic Grading of Barley Grain for Brewery Industries Using Convolutional Neural Network. Available at SSRN 5064645.

Text ©. The authors.
Except where otherwise noted, content and images are subject to copyright. Any reuse without express permission from the copyright owner is prohibited.