---
license: apache-2.0
tags:
  - image-classification
  - x-ray
  - customs
  - smuggling-detection
  - sonar
language:
  - ar
  - en
---

# 🔱 SONAR AI - Customs X-Ray Inspection Models

## Dr. Abbas Fadel Jassim | 2026

> AI-powered customs inspection system for X-ray cargo scanning

---

## 📊 Models Performance Summary

| # | Model | Task | Classes | F1 Score | Architecture |
|---|-------|------|---------|----------|-------------|
| 1 | **Classification** | تصنيف البضاعة | 43 | **94.2%** | Swin-V2 Tiny |
| 2 | **Concealment** | كشف الإخفاء | 2 | **98.4%** | Swin-V2 Tiny |
| 3 | **Risk Assessment** | تقييم الخطورة | 5 | **97.3%** | DaViT Tiny |

---

## 🔍 Model 1: Classification (43 cargo categories)

**Task:** Classify X-ray images into 43 different cargo types

| Model | Params | Accuracy | F1 Score |
|-------|--------|----------|----------|
| 🏆 **Swin-V2 Tiny** | 27.6M | 94.2% | **94.1%** |
| 🥈 **DaViT Tiny** | 27.6M | 94.2% | **94.1%** |

### Per-Class Performance (Swin-V2):

| Category | Precision | Recall | F1 | Samples |
|----------|-----------|--------|-----|---------|
| appliances | 98.2% | 98.2% | **98.2%** | 57 |
| auto_parts | 92.6% | 94.3% | **93.5%** | 53 |
| bags | 100% | 100% | **100%** | 12 |
| banana | 100% | 100% | **100%** | 40 |
| batteries | 100% | 88.9% | **94.1%** | 9 |
| beverages | 71.4% | 100% | **83.3%** | 10 |
| cables | 100% | 94.4% | **97.1%** | 18 |
| canned_food | 92.0% | 79.3% | **85.2%** | 29 |
| ceramic | 97.5% | 97.5% | **97.5%** | 40 |
| chemicals | 100% | 94.9% | **97.4%** | 39 |
| cleaning | 91.7% | 91.7% | **91.7%** | 24 |
| clothes | 91.8% | 96.3% | **94.0%** | 81 |
| cooking_oil | 76.9% | 100% | **87.0%** | 10 |
| cosmetics | 100% | 94.1% | **97.0%** | 17 |
| electronics | 92.7% | 87.9% | **90.3%** | 58 |
| fruits | 85.7% | 100% | **92.3%** | 6 |
| furniture | 97.1% | 96.6% | **96.8%** | 174 |
| glass | 95.2% | 100% | **97.6%** | 20 |
| kitchenware | 100% | 100% | **100%** | 2 |
| lubricants | 92.3% | 100% | **96.0%** | 12 |
| machinery | 87.5% | 95.5% | **91.3%** | 66 |
| meat | 92.3% | 92.3% | **92.3%** | 13 |
| medical | 100% | 100% | **100%** | 3 |
| milk | 98.4% | 98.4% | **98.4%** | 62 |
| motorcycle | 100% | 66.7% | **80.0%** | 6 |
| nuts | 92.9% | 100% | **96.3%** | 26 |

**Training:** 30 epochs, AdamW (lr=1e-4), CosineAnnealing, Albumentations augmentation

---

## 🔍 Model 2: Concealment Detection (match vs no_match)

**Task:** Detect hidden/smuggled items in X-ray containers

| Model | Params | Accuracy | F1 Score |
|-------|--------|----------|----------|
| 🏆 **Swin-V2 Tiny** | 27.6M | 98.4% | **98.4%** |
| 🥈 **DaViT Tiny** | 27.6M | 98.4% | **98.4%** |
| 🥉 **EVA-02 Tiny** | 5.5M | 97.8% | **97.8%** |
| 4️⃣ **MaxViT Tiny** | 30.4M | 96.7% | **96.8%** |

### Classes:
- **match:** Container contents match declaration (6,348 images)
- **no_match:** Concealed/smuggled items detected (702 images)

---

## ⚠️ Model 3: Risk Assessment (5 levels)

**Task:** Assess risk level of container cargo

| Model | Params | Accuracy | F1 Score |
|-------|--------|----------|----------|
| 🏆 **DaViT Tiny** | 27.6M | 97.2% | **97.3%** |
| 🥈 **Swin-V2 Tiny** | 27.6M | 97.2% | **97.2%** |

### Risk Levels:

| Level | Name (AR) | Name (EN) | Samples | F1 Score |
|-------|-----------|-----------|---------|----------|
| 0 | آمن | Safe | 6,355 | **99.1%** |
| 1-2 | منخفض | Low | 5 | ⚠️ Limited |
| 3 | متوسط | Medium | 387 | **79.5%** |
| 4 | عالي | High | 297 | **84.5%** |
| 5 | حرج | Critical | 6 | ⚠️ Limited |

**Note:** Low and Critical levels have very few training samples.

---

## 📁 Files
```
SONAR-AI-Models/
├── concealment/
│   ├── best_swinv2.pth      (110 MB, 98.4% F1)
│   ├── best_davit.pth       (110 MB, 98.4% F1)
│   ├── best_eva02.pth       (22 MB, 97.8% F1)
│   └── best_maxvit.pth      (122 MB, 96.8% F1)
├── classification/
│   ├── best_swinv2_43cls.pth (110 MB, 94.1% F1)
│   └── best_davit_43cls.pth  (110 MB, 94.1% F1)
├── risk/
│   ├── best_swinv2_risk.pth  (110 MB, 97.2% F1)
│   └── best_davit_risk.pth   (110 MB, 97.3% F1)
└── results/
    ├── classification_cm.png
    ├── classification_training.png
    ├── risk_cm.png
    └── risk_training.png
```

## 🛠️ Usage
```python
import timm, torch
from torchvision import transforms
from PIL import Image

# Load model
model = timm.create_model('swinv2_tiny_window8_256', pretrained=False, num_classes=43)
ckpt = torch.load('classification/best_swinv2_43cls.pth', map_location='cpu')
model.load_state_dict(ckpt['model_state_dict'])
model.eval()

# Inference
tf = transforms.Compose([
    transforms.Resize((256, 256)),
    transforms.ToTensor(),
    transforms.Normalize([0.485,0.456,0.406], [0.229,0.224,0.225])
])
img = tf(Image.open('xray.jpg').convert('RGB')).unsqueeze(0)
with torch.no_grad():
    pred = model(img).argmax(1).item()
```

## 📜 License
Apache 2.0

## 👨‍💻 Author
**Dr. Abbas Fadel Jassim** - 2026