20-Fruit_Recognition

Fruit Recognition System

ResNet18 + CBAM Attention · 20-Class Fruit Classification · Test Accuracy: 91.78%

Click here to the repo of Github

---

Model Architecture

Input (224×224 RGB)
        │
        ▼
┌─────────────────────────────────┐
│  ResNet18 Backbone (ImageNet)   │
│  ├─ Conv1 + BN + ReLU + MaxPool │
│  ├─ Layer1 (64ch)               │
│  ├─ Layer2 (128ch) ──► CBAM(128)│
│  ├─ Layer3 (256ch) ──► CBAM(256)│
│  └─ Layer4 (512ch) ──► CBAM(512)│
│     └─ AdaptiveAvgPool2d(1,1)   │
└─────────────────────────────────┘
        │ 512-d feature
        ▼
┌─────────────────────────────────┐
│  Classification Head            │
│  Dropout(0.5)                   │
│  Linear(512→256) + ReLU + BN    │
│  Dropout(0.3)                   │
│  Linear(256→20)                 │
└─────────────────────────────────┘
        │
   20-class logits

Total Parameters: ~11.39M

---

Key Techniques

CBAM Attention

Channel + Spatial dual-dimension attention, injected after ResNet Layer2/3/4. Guides the model to focus on fruit regions and suppress background interference.

Two-Stage Progressive Training

• Phase 1 (10 epochs): Freeze backbone, train CBAM + classifier head (lr=1e-3). Only ~213K trainable parameters.
• Phase 2 (15 epochs): Unfreeze Layer3/4, differential learning rates (backbone=1e-4, head=5e-4).

CutMix Augmentation

Mixed-sample augmentation with probability 0.5. Forces the model to learn from partial regions, improving generalization.

Class Imbalance Handling

Dual mechanism: weighted random sampling + weighted cross-entropy loss to ensure minority class performance.

---

20 Fruit Classes

---

References

• ResNet: He et al., Deep Residual Learning for Image Recognition, 2016
• CBAM: Woo et al., Convolutional Block Attention Module, 2018
• CutMix: Yun et al., CutMix: Regularization Strategy, 2019