FDSNet

FDSNet: An Accurate Real-Time Surface Defect Segmentation Network - ICASSP 2022 paper [pdf]

Dataset

⭐️ MSD dataset ⭐️

The generated auxiliary ground-truth AuxiliaryGT for MSD dataset. The images of MSD dataset are downsampled to 1440×810 during training and test.
We convert SD-saliency-900 and Magnetic-tile-defect-datasets (denoted as MT-Defect) dataset to PASCAL VOC format and divide the datasets into train: val: test = 6: 2: 2 randomly. We use trainval-test for NEU-Seg and MT-Defect and train-test for MSD dataset. The converted datasets can be downloaded here: MT-Defect and NEU-Seg.

Environment

Python 3.8.5 PyTorch 1.9.0 CUDA 11.1
one NVIDIA GTX 1080Ti GPU

conda env create -f requirements.yml

Usage

First download the dataset and the auxiliary ground-truth. Put the auxiliary GT to the data folder and modify the path in the /core/data/dataloader.
when train model on NEU-Seg, set scale-ratio=None. when train model on MT-Defect, set crop size=450 and base_size not None.
Train model

CUDA_VISIBLE_DEVICES=0 python train.py --model fdsnet --use-ohem True --aux True --dataset phone_voc --scale-ratio 0.75 --lr 0.0001 --epochs 150 --batch-size 8

Eval model. We eval the image one by one.

python eval.py

Pretrained Model

Dataset	Pth	mIoU	FPS
MSD	fastscnn__phone_voc_best_model.pth	89.1	115.0
MSD	fdsnet__phone_voc_best_model.pth	90.2	135.0
MT-Defect	fdsnet__mt_voc_best_model.pth	63.9	181.5
NEU-Seg	fdsnet__sd_voc_best_model.pth	78.8	186.1

Results

Code Borrow

Semantic Segmentation on PyTorch
Fast-SCNN