SegFormer model is free and open-source! Fine-tuned based on a specific architecture, it accurately detects defects in various materials.

Segformer Finetuned 4ss1st3r S3gs3m 24Jan All 10k Steps

Developed by blzncz

An image segmentation model fine-tuned based on nvidia/mit-b0 architecture, focused on detecting various types of material defects

Image Segmentation

Transformers

Open Source License:Other #Industrial defect detection #High-precision segmentation #Material surface analysis

Downloads 16

Release Time : 2/7/2024

Model Overview

This model is a fine-tuned SegFormer variant on the blzncz/4ss1st3r_s3gs3m_24Jan_all dataset, specifically designed for industrial material surface defect detection, capable of identifying various defect types such as cohesive defects, mesh defects, adhesive defects, and bubble defects

Model Features

Multi-defect type detection

Capable of simultaneously detecting various industrial material defects such as cohesive defects, mesh defects, adhesive defects, and bubble defects

High-precision segmentation

Achieves 96.68% accuracy and 91.67% intersection over union (IoU) in cohesive defect detection

Industrial scenario optimization

Optimized specifically for industrial material surface defect detection scenarios, maintaining good performance even in complex backgrounds

Model Capabilities

Image segmentation

Defect detection

Industrial quality inspection

Material surface analysis

Use Cases

Industrial quality inspection

Material surface defect detection

Automated detection of various defect types on industrial material surfaces

96.68% accuracy in cohesive defect detection, 68.08% accuracy in mesh defect detection

Production line quality monitoring

Integrated into production lines for real-time product quality monitoring

Overall accuracy reaches 92.60%

🚀 segformer-finetuned-4ss1st3r_s3gs3m_24Jan_all-10k-steps

This model is a fine - tuned version of nvidia/mit-b0 on the blzncz/4ss1st3r_s3gs3m_24Jan_all dataset. It is designed for image segmentation tasks in the vision field, achieving excellent results on the evaluation set.

🚀 Quick Start

This model is ready to use for image segmentation tasks. You can fine - tune it further according to your specific needs.

✨ Features

Fine - tuned: Based on nvidia/mit-b0, it has been fine - tuned on the blzncz/4ss1st3r_s3gs3m_24Jan_all dataset.
Good Performance: Achieves the following results on the evaluation set:
- Loss: 0.3095
- Mean Iou: 0.5513
- Mean Accuracy: 0.7874
- Overall Accuracy: 0.9260
- Accuracy Bg: nan
- Accuracy Fallo cohesivo: 0.9668
- Accuracy Fallo malla: 0.6808
- Accuracy Fallo adhesivo: 0.9727
- Accuracy Fallo burbuja: 0.5291
- Iou Bg: 0.0
- Iou Fallo cohesivo: 0.9167
- Iou Fallo malla: 0.6189
- Iou Fallo adhesivo: 0.7307
- Iou Fallo burbuja: 0.4903

📚 Documentation

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 6e - 05
train_batch_size: 8
eval_batch_size: 8
seed: 1337
optimizer: Adam with betas=(0.9,0.999) and epsilon = 1e - 08
lr_scheduler_type: polynomial
training_steps: 10000

Training results

Training Loss	Epoch	Step	Validation Loss	Mean Iou	Mean Accuracy	Overall Accuracy	Accuracy Bg	Accuracy Fallo cohesivo	Accuracy Fallo malla	Accuracy Fallo adhesivo	Accuracy Fallo burbuja	Iou Fallo cohesivo	Iou Fallo malla	Iou Fallo adhesivo	Iou Fallo burbuja
0.1378	1.0	783	0.2677	0.4895	0.7143	0.9122	nan	0.9724	0.5531	0.9663	0.3654	0.9038	0.5327	0.6757	0.3351
0.1117	2.0	1566	0.2305	0.5289	0.7978	0.9246	nan	0.9507	0.7727	0.9705	0.4974	0.9214	0.6808	0.5876	0.4549
0.0881	3.0	2349	0.2041	0.5556	0.7867	0.9354	nan	0.9712	0.7391	0.9389	0.4975	0.9273	0.6790	0.7323	0.4394
0.0878	4.0	3132	0.1984	0.5584	0.8003	0.9346	nan	0.9556	0.8247	0.9602	0.4606	0.9261	0.6935	0.7373	0.4352
0.0895	5.0	3915	0.2841	0.5246	0.8086	0.9088	nan	0.9137	0.8834	0.9719	0.4652	0.8964	0.6309	0.6593	0.4365
0.0773	6.0	4698	0.2547	0.5652	0.7823	0.9336	nan	0.9775	0.6843	0.9384	0.5291	0.9251	0.6378	0.7820	0.4813
0.0667	7.0	5481	0.2726	0.5609	0.7932	0.9295	nan	0.9741	0.6609	0.9689	0.5689	0.9203	0.6202	0.7548	0.5093
0.0678	8.0	6264	0.2950	0.5276	0.8002	0.9175	nan	0.9443	0.7561	0.9713	0.5292	0.9089	0.6570	0.5900	0.4822
0.0653	9.0	7047	0.2712	0.5467	0.7682	0.9288	nan	0.9690	0.6971	0.9641	0.4425	0.9189	0.6330	0.7588	0.4228
0.0646	10.0	7830	0.2841	0.5499	0.7819	0.9272	nan	0.9681	0.6840	0.9688	0.5068	0.9178	0.6243	0.7345	0.4728
0.057	11.0	8613	0.3373	0.5257	0.7782	0.9166	nan	0.9593	0.6555	0.9739	0.5242	0.9075	0.6040	0.6319	0.4848
0.0591	12.0	9396	0.3082	0.5504	0.7900	0.9247	nan	0.9656	0.6776	0.9705	0.5463	0.9148	0.6172	0.7182	0.5019
0.053	12.77	10000	0.3095	0.5513	0.7874	0.9260	nan	0.9668	0.6808	0.9727	0.5291	0.9167	0.6189	0.7307	0.4903

Framework versions

Transformers 4.31.0.dev0
Pytorch 2.0.1+cpu
Datasets 2.13.1
Tokenizers 0.13.3

📄 License

This model is released under the "other" license.

Featured Recommended AI Models

Empowering the Future, Your AI Solution Knowledge Base

English 简体中文繁體中文にほんご