Damage Level Classification from Images Using Deep Learning with Data Augmentation

0. Introduction

This project is dedicated to explore the best cost-performance value of deploying ResNet model in different sizes of database.
Accuracy = correct/total

1. result

Model(833 images)	Val Loss	Val Acc	Test Acc	Training Time	Epoch Ended	Rank
ResNet18	0.9694	0.7246	0.7289	60s	13	5
ResNet18 + fine-tuning	0.5266	0.8024	0.7711	60s	18	3
ResNet18 + data Augmentation	0.7036	0.7365	0.7229	120s	34	6
ResNet18 + fine-tuning + Augmentation	0.5409	0.7844	0.7711	170s	47	4
ResNet50	0.8773	0.6587	0.6145	70s	4	8
ResNet50 + fine-tuning	0.6015	0.8084	0.7771	100s	9	2
ResNet50 + data Augmentation	0.2832	0.6527	0.6386	140s	18	7
ResNet50 + fine-tuning + Augmentation	0.6202	0.7725	0.8012	140s	16	1

Model(2280 images)	Val Loss	Val Acc	Test Acc	Training Time	Epoch Ended	Rank
ResNet18	0.4777	0.8509	0.8399	80s	6	6
ResNet18 + fine-tuning	0.3322	0.8925	0.8772	120s	10	3
ResNet18 + data Augmentation	0.4773	0.8531	0.8553	240s	29	5
ResNet18 + fine-tuning + Augmentation	0.3962	0.8772	0.8662	720s	51	4
ResNet50	0.5105	0.818	0.8158	150s	8	8
ResNet50 + fine-tuning	0.3338	0.8772	0.8904	180s	6	2
ResNet50 + data Augmentation	0.5119	0.8311	0.8311	450s	33	7
ResNet50 + fine-tuning + Augmentation	0.3888	0.8816	0.8947	300s	18	1

Model(19104 images)	Val Loss	Val Acc	Test Acc	Training Time	Epoch Ended	Rank
ResNet18	0.5387	0.8129	0.8223	400s	5	6
ResNet18 + fine-tuning	0.4644	0.8411	0.8427	600s	6	4
ResNet18 + data Augmentation	0.5424	0.8076	0.7995	840s	14	8
ResNet18 + fine-tuning + Augmentation	0.4349	0.839	0.8487	1600s	34	1
ResNet50	0.5461	0.8076	0.8154	1200s	7	5
ResNet50 + fine-tuning	0.4291	0.8396	0.8448	780s	3	3
ResNet50 + data Augmentation	0.5306	0.8074	0.8026	3100s	38	7
ResNet50 + fine-tuning + Augmentation	0.4401	0.8469	0.8463	1800s	18	2

2. Takeaway

1. Why use data augmentation?

• the model sees the same base dataset, but different augmented images in every epoch. Even if the dataset is small, augmentation effectively creates "new" training samples by introducing variations, making the model more robust.

2. While using early stop, Val loss or Val acc?

• Using Val loss is better, espeically when the class is not balanced. If using Val acc, the model would prefer to guess the dominant class. In other words, the model would lost the ability to forceast the rare case.

3. Epoch & Patience

• Patience should be 10–20% of the total epochs.

3. ResNet18 v.s. ResNet50?

Feature	ResNet-18	ResNet-50
Architecture Depth	18 layers	50 layers
Parameters	~11.7M	~25.6M
Training Speed	Faster (lighter)	Slower (heavier)
Inference Speed	Faster	Slower
Model Size	Smaller	Larger
Capacity	Lower (simpler tasks)	Higher (complex tasks)