Comprehensive Report: Comparison of Image Classification Models

1. Model Overview

This report summarizes and compares experimental results of four different image classification models:

1. EfficientNet
2. ResNet34
3. Vision Transformer (ViT)
4. YOLOv8n

2. Training Configuration Comparison

Parameter	EfficientNet	ResNet34	Vision Transformer	YOLOv8n
Number of epochs	15	100	100	10
Learning rate	0.001 (assumed)	0.1→0.01→0.001→0.0001	0.00001 (fixed)	Increase-decrease schedule
Special techniques	Early stopping (patience=5)	Learning rate schedule	None	Learning rate schedule

3. Experimental Results Comparison

3.1. Highest Accuracy

Model	Highest Accuracy	On Dataset	Timing
EfficientNet	97.04%	Test	Epoch 11 (best)
ResNet34	80.74%	Test	After 100 epochs
Vision Transformer	74.81%	Test	After 100 epochs
YOLOv8n	95.9%	Test	Epoch 10

3.2. Loss Evolution

Model	Initial Loss	Final Loss	Characteristics
EfficientNet	0.0717	0.0583	Irregular fluctuations from 0.0330-0.1081
ResNet34	4.41	0.4341	Decreases in LR phases, stable at end
Vision Transformer	1.073	0.0004	Steady decrease, converges near 0
YOLOv8n	0.83442	0.07431	Fast decrease and stable

3.3. Training Time and Performance

Model	Training Time	Inference Performance	Fast Convergence
EfficientNet	Not specified	Not specified	Achieves >92% from first epoch
ResNet34	Not specified	Not specified	Slow loss decrease, stable
Vision Transformer	Not specified	Not specified	Steady loss decrease, final convergence
YOLOv8n	Short (10 epochs)	17.4ms/image	Achieves >90% after just 2 epochs

4. Detailed Class-wise Analysis

4.1. EfficientNet Class-wise Metrics

Class	Precision	Recall	F1-score
Adult	1.0000	1.0000	1.0000
Normal	0.9528	0.9712	0.9619
Violent	0.9595	0.9342	0.9467

4.2. EfficientNet Confusion Matrix

                        Predicted
                adult   normal violence
True    Adult       90        0        0
True   Normal        0      101        3
True  Violent        0        5       71

4.3. Other Models Metrics

• ResNet34: Only reports overall accuracy (80.74%), no detailed class-wise metrics
• Vision Transformer: Only reports overall accuracy (74.81%), no detailed class-wise metrics
• YOLOv8n: Reports Top-1 Accuracy (95.9%) and Top-5 Accuracy (100%)

5. Feature and Advantage Comparison

5.1. EfficientNet

• Advantages: Highest accuracy (97.04%), perfect classification for "Adult" class
• Characteristics: Validation accuracy fluctuates significantly, loss doesn't decrease evenly

5.2. ResNet34

• Advantages: Effective learning rate schedule
• Characteristics: Stable training, no sudden loss fluctuations

5.3. Vision Transformer

• Advantages: Training loss decreases to near 0 (0.0004)
• Characteristics: Low performance (74.81%) suggests possible overfitting

5.4. YOLOv8n

• Advantages: Fast convergence, high inference speed (17.4ms/image)
• Characteristics: High efficiency with few epochs (10)

6. Overall Performance Comparison

Criteria	EfficientNet	ResNet34	Vision Transformer	YOLOv8n
Accuracy	★★★★★ (97.04%)	★★★☆☆ (80.74%)	★★☆☆☆ (74.81%)	★★★★★ (95.9%)
Convergence speed	★★★★☆ (15 epochs)	★★☆☆☆ (100 epochs)	★★☆☆☆ (100 epochs)	★★★★★ (10 epochs)
Loss stability	★★☆☆☆ (fluctuating)	★★★★☆ (stable decrease)	★★★★★ (steady decrease)	★★★★☆ (fast decrease)
Inference time	No information	No information	No information	★★★★★ (17.4ms/image)
Overall rating	★★★★★	★★★☆☆	★★☆☆☆	★★★★★

7. Conclusions and Recommendations

7.1. General Conclusions

From the experimental results of all four models, the following conclusions can be drawn:

1. EfficientNet and YOLOv8n are the two most effective models with >95% accuracy on test set, suitable for content-based image classification tasks.

2. ResNet34 has moderate performance (80.74%) but stable training process, may need additional optimization.

3. Vision Transformer has the lowest performance (74.81%) on the MNIST dataset, suggesting that transformer architecture may not be suitable for simple data or needs hyperparameter adjustments.

4. YOLOv8n stands out with fast convergence capability (only 10 epochs) and high inference speed, suitable for practical applications and real-time processing.

7.2. Improvement Recommendations

For EfficientNet:

• Enhance data for "Normal" and "Violent" classes to reduce confusion
• Apply learning rate scheduler to stabilize training process

For ResNet34:

• Optimize learning rate schedule
• Apply additional regularization techniques (dropout, weight decay)
• Experiment with larger architectures (ResNet50, ResNet101)

For Vision Transformer:

• Apply regularization to avoid overfitting
• Experiment with learning rate schedule instead of fixed rate
• Adjust hyperparameters (number of layers, attention heads, patch size)

For YOLOv8n:

• Optimize dataset (increase number of hard samples)
• Adjust hyperparameters to reduce val loss fluctuation

7.3. Model Selection by Problem Type

• Need highest accuracy: EfficientNet (97.04%)
• Need fast processing speed: YOLOv8n (17.4ms/image)
• Balanced practical application: YOLOv8n (95.9% accuracy, 10 epochs, fast inference)
• Complex data with many details: ResNet34 with learning rate schedule
• Complex spatial structure data: Vision Transformer (after optimization)

8. Individual Model Reports

1. Analysis report of EfficientNet model for three-label image classification
2. Experimental results report of ResNet34 model
3. Experimental results report of Vision Transformer model
4. Performance analysis report of 16+ image classification project using YOLOv8n model

If you find any errors or areas for improvement, please don't hesitate to add an issue so the AI community can continue to grow :))