Image-Based Classification of Olive Varieties Native to Turkiye Using Multiple Deep Learning Architectures: Analysis of Performance, Complexity, and Generalization

This study evaluates ten deep learning architectures for classifying five native Turkish black table olive varieties using a 2,500-image dataset, finding that EfficientNetV2-S achieved the highest accuracy (95.8%) while EfficientNetB0 offered the optimal balance between performance and computational efficiency, ultimately demonstrating that parametric efficiency is more critical than model depth under limited data conditions.

The Gist

Imagine you are a master olive farmer in Türkiye. You have five different types of black olives: Gemlik, Ayvalık, Uslu, Erkence, and Çelebi. To the untrained eye, they look almost identical. They are all round, dark, and shiny. But to a connoisseur, they are as different as a Ferrari is from a Ford.

Sorting them by hand is slow, tiring, and prone to human error. So, the researchers in this paper asked: "Can we teach a computer to sort these olives perfectly?"

To find the answer, they didn't just pick one "smart" computer program. Instead, they held a 10-way race between ten different types of Artificial Intelligence (AI) brains to see which one was the best at this specific job.

Here is the story of that race, explained simply.

The Contestants: The AI Athletes

The researchers lined up ten different AI models. Think of them as athletes with different body types and training styles:

1. The Heavyweights (Deep CNNs): Models like ResNet and DenseNet. These are like bodybuilders. They are huge, have massive muscles (parameters), and can lift heavy loads. They are powerful but slow and hungry for food (data).
2. The Sprinters (EfficientNet & MobileNet): Models like EfficientNet and MobileNet. These are like Olympic sprinters. They are lean, fast, and incredibly efficient. They do a lot with very little energy.
3. The Visionaries (Transformers): Models like ViT and Swin. These are the new kids on the block. Instead of looking at an image piece-by-piece like a human, they try to understand the "whole picture" at once using a complex attention mechanism. They are brilliant but usually need a massive library of books (data) to learn.

The Training Ground: The Olive Gym

The researchers built a gym for these athletes. They gathered 2,500 photos of the five olive types (500 of each).

• The Setup: They took photos in a clean, white room with perfect lighting so the olives looked their best.
• The Rules: They split the photos into three groups: one for learning (training), one for checking homework (validation), and one final exam (testing) that the models had never seen before.

They taught the models using a technique called Transfer Learning. Imagine you are teaching a chef who already knows how to cook French cuisine how to cook Turkish food. You don't start from scratch; you just teach them the new recipes. Similarly, these AI models were already "smart" from learning on millions of other images (like cats and cars), and the researchers just fine-tuned them to recognize olives.

The Race Results: Who Won?

🏆 The Gold Medalist: EfficientNetV2-S

This model was the fastest and most accurate. It got 95.8% of the olives right.

• The Analogy: It was like a master chef who could taste a dish and instantly know exactly which spices were used. It saw the tiny differences in the olive's shape and skin texture that humans might miss.

🥈 The Silver Medalist (The Smartest Choice): EfficientNetB0

This model got 94.5% right. It wasn't quite as accurate as the Gold Medalist, but here is the kicker: it was 20 times more efficient.

• The Analogy: Imagine two cars. The Gold Medalist is a Formula 1 car: fast and amazing, but it guzzles gas and costs a fortune to maintain. The Silver Medalist is a high-end hybrid: slightly slower, but it gets incredible mileage and is cheap to run. For a real-world factory, the hybrid is often the better buy.

🥉 The Underperformers: The Heavyweights & Visionaries

• The Heavyweights (ResNet/DenseNet): They did well, but they were "overkill." They used too much energy for the job.
• The Visionaries (ViT-B16): This was the biggest surprise. The most complex model, which usually wins big competitions, actually lost. It only got 88.5% right and confused the olives a lot.
• Why? The researchers realized that the Visionary models are like a student who tries to read an entire encyclopedia to learn how to tie a shoelace. They need massive amounts of data to work well. With only 2,500 olives, they got confused and started "memorizing" the answers instead of learning the rules (a problem called overfitting).

The Big Lesson: Bigger Isn't Always Better

The most important takeaway from this paper is a simple rule: In the world of AI, bigger brains don't always mean smarter results.

When you have a small dataset (like a limited number of olive photos), a lean, efficient model works better than a giant, complex one. The complex models got confused because they had too many "neurons" for the amount of information they were given.

What Does This Mean for the Real World?

The researchers aren't just writing a paper; they are solving a real problem.

• For a small farm with a cheap phone: Use the MobileNet model. It's light enough to run on a phone and fast enough to sort olives on a conveyor belt.
• For a big factory with powerful computers: Use EfficientNetV2-S for the absolute best accuracy.
• For the best balance: Use EfficientNetB0. It gives you 95% accuracy without needing a supercomputer.

Summary

This study is like a test drive for ten different cars to see which one is best for a specific trip. They found that you don't need a rocket ship to drive to the grocery store; a well-tuned, efficient car gets you there faster, cheaper, and with less trouble.

The Bottom Line: To sort Turkish olives, you don't need the biggest, most complex AI. You need the right-sized, efficient AI that knows how to look at the details without getting overwhelmed.

Technical Summary

1. Problem Statement

The study addresses the challenge of fine-grained classification for five locally cultivated black table olive varieties in Türkiye: Gemlik, Ayvalık, Uslu, Erkence, and Çelebi.

• Context: These varieties are morphologically similar (differing subtly in size, shape, skin color, and texture), making manual visual inspection subjective, error-prone, and inefficient for high-throughput agricultural quality control.
• Gap: While deep learning has been applied to agricultural sorting, there is a lack of systematic comparative analyses focusing on Türkiye-specific varieties using modern architectures (including Transformers and EfficientNets) under limited data conditions.
• Core Question: Does increasing model depth/parameters guarantee better accuracy in small datasets, and which architecture offers the best trade-off between performance and computational cost?

2. Methodology

The researchers employed a rigorous experimental framework involving dataset construction, preprocessing, and a comparative evaluation of ten deep learning models.

A. Dataset Construction

• Source: 2,500 images of black table olives captured in a controlled lab environment with uniform lighting and a matte white background.
• Classes: 5 varieties (500 images each), ensuring a fully balanced dataset.
• Split: Stratified random split into Training (80%), Validation (10%), and Test (10%). The test set contained 50 images per class.

B. Preprocessing & Augmentation

• Standardization: Images resized to 224×224 pixels.
• Noise Reduction: Gaussian smoothing applied.
• Segmentation: Otsu's thresholding used to create binary masks, isolating the olive from the background to focus learning on morphological features.
• Augmentation: Applied only to the training set (random rotation, horizontal flip, brightness jitter, scaling) to prevent overfitting.

C. Architectures Evaluated

Ten models were trained using Transfer Learning (ImageNet pre-trained weights) with a 5-class softmax output:

1. Lightweight CNNs: MobileNetV2, EfficientNetB0.
2. Efficient Variants: EfficientNetV2-S.
3. Deep CNNs: ResNet50, ResNet101, DenseNet121, InceptionV3.
4. Modern CNN: ConvNeXt-Tiny.
5. Transformers: Vision Transformer (ViT-B16), Swin Transformer (Swin-T).

D. Training & Evaluation

• Training: Adam optimizer, learning rate 1e-3, batch size 32, max 25 epochs with early stopping.
• Metrics:
  • Performance: Accuracy, Precision, Recall, F1-score, MCC, Cohen's Kappa, ROC-AUC.
  • Generalization: Generalization gap (Training Accuracy – Validation Accuracy).
  • Complexity: Parameter count, FLOPs (Floating Point Operations), inference time, and accuracy-to-FLOPs ratio.

3. Key Contributions

1. Dataset Creation: A balanced, high-quality dataset of five specific Turkish olive varieties designed for fine-grained classification.
2. Comprehensive Benchmarking: A unified comparison of 10 diverse architectures (CNNs vs. Transformers) under identical experimental conditions.
3. Holistic Evaluation: Moving beyond simple accuracy to include statistical significance testing, generalization gap analysis, and computational efficiency metrics (FLOPs/parameters).
4. Practical Guidelines: Specific recommendations for model selection based on deployment constraints (e.g., edge devices vs. centralized servers).

4. Key Results

A. Performance Rankings

• Highest Accuracy: EfficientNetV2-S achieved 95.8% accuracy with an MCC and Kappa of 0.95.
• Runner-up: EfficientNetB0 achieved 94.5% accuracy.
• Other Performers: ConvNeXt-Tiny and Swin-T (~94%); MobileNetV2 (92.8%).
• Lowest Performance: ViT-B16 achieved only 88.5%, despite having the highest parameter count (~86M).

B. Generalization & Overfitting

• EfficientNetV2-S: Minimal generalization gap (1.6%), indicating stable learning.
• ViT-B16: Severe overfitting with a 10% generalization gap (98.5% train vs. 88.5% val). This suggests Transformer architectures are less data-efficient for small, fine-grained datasets compared to CNNs.
• Conclusion: Under limited data, parametric efficiency is more critical than raw model depth or parameter count.

C. Computational Efficiency

• Best Trade-off: EfficientNetB0 offered the best balance, delivering 94.5% accuracy with only 0.39 GFLOPs.
• Edge Suitability: MobileNetV2 (0.30 GFLOPs) is ideal for embedded systems, achieving 92.8% accuracy.
• Inefficiency: ViT-B16 required 17.6 GFLOPs for the lowest accuracy, making it computationally disadvantageous for this task.

D. Error Analysis

• The highest confusion occurred between Erkence and Çelebi due to similar fruit diameter and skin color.
• Uslu was the easiest to classify due to distinctive morphological traits.

5. Significance and Implications

• Paradigm Shift: The study challenges the assumption that "larger models are always better." It demonstrates that for small, balanced agricultural datasets, optimized, moderately sized CNNs outperform highly parameterized Transformers.
• Deployment Strategy:
  • Edge Devices: Use MobileNetV2.
  • Balanced Systems: Use EfficientNetB0.
  • High-Accuracy Centralized Systems: Use EfficientNetV2-S.
• Future Directions: The authors suggest that future improvements could involve explicit morphological feature extraction or hyperspectral imaging to resolve the confusion between visually similar varieties like Erkence and Çelebi.

Final Conclusion: The research establishes that for fine-grained agricultural classification with limited data, parametric efficiency and architectural optimization (specifically the EfficientNet family) are superior to simply increasing model depth or adopting Transformer-based architectures.