BiDNet

This repo is the official implementation of "BiDNet: A Real-Time Semantic Segmentation Network with Anti-Feature Interference and Detail Recovery for Industrial Defects". This repo contains the supported code, configuration files, and datasets to reproduce the semantic segmentation results of BiDNet. The code is mainly based on MMSegmentaion V1.2.2. All experiments were performed on a single NVIDIA GTX 3090Ti GPU in CUDA 11.7, Python 3.8, and PyTorch 1.13.1.

Code Snippet

The code snippet is here.

Citation

If you find our repo useful for your research, please consider citing our paper: @ARTICLE{BiDNet, author={Pan, Jiawei and Zeng, Deyu and Wu, Zongze and Xie, Shengli}, journal={IEEE Transactions on Instrumentation and Measurement}, title={BiDNet: A Real-Time Semantic Segmentation Network with Anti-Feature Interference and Detail Recovery for Industrial Defects}, year={2025}, volume={74}, number={}, pages={1-16}, doi={10.1109/TIM.2025.3548182}}

Installation

Step 1. Create a conda environment and activate it. conda create --name BiDNet python=3.8 -y conda activate BiDNet Step 2. Install PyTorch following official instructions, e.g. conda install pytorch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 pytorch-cuda=11.7 -c pytorch -c nvidia -y Step 3. Install MMCV using MIM. pip install -U openmim mim install mmengine mim install mmcv==2.0.0rc4 Step 4. Install BiDNet. git clone -b main https://github.com/jiaweipan997/BiDNet.git cd BiDNet pip install -v -e . pip install ftfy pip install regex

Verify the installation

Step 1. We need to download config and checkpoint files. When it is done, you will find two files pspnetr50-d84xb2-40kcityscapes-512x1024.py and pspnetr50-d8512x102440kcityscapes20200605003338-2966598c.pth in your current folder. ``` mim download mmsegmentation --config pspnetr50-d84xb2-40kcityscapes-512x1024 --dest . Step 2. Verify the inference demo. You will see a new image result.jpg on your current folder, where segmentation masks are covered on all objects. python demo/imagedemo.py demo/demo.png configs/pspnet/pspnetr50-d84xb2-40kcityscapes-512x1024.py pspnetr50-d8512x102440kcityscapes20200605003338-2966598c.pth --device cuda:0 --out-file result.jpg ```

Dataset prepare

The preprocessed dataset can be downloaded from this link with the code s4hn. Please create a new data folder and put the downloaded dataset in it and unzip it. The structure is as follows： BiDNet ├── mmseg ├── tools ├── configs ├── data │ ├── MSD │ │ ├── imgs │ │ │ ├── train │ │ │ ├── val │ │ │ ├── test │ │ ├── labels │ │ │ ├── train │ │ │ ├── val │ │ │ ├── test . . .

Training

```

MSD dataset:

bash tools/disttrain.sh configs/BiDNet/bidnetmscan-t1xb8-adamw-80kmsd-512x512.py 1

MTD dataset:

bash tools/disttrain.sh configs/BiDNet/bidnetmscan-t1xb4-adamw-40kmtd-512x512.py 1

GSD dataset:

bash tools/disttrain.sh configs/BiDNet/bidnetmscan-t1xb16-adamw-160kgsd-512x512.py 1 ```

Evaluation

```

MSD dataset:

python tools/test.py configs/BiDNet/bidnetmscan-t1xb8-adamw-80kmsd-512x512.py workdirs/bidnetmscan-t1xb8-adamw-80kmsd-512x512/iter80000.pth

MTD dataset:

python tools/test.py configs/BiDNet/bidnetmscan-t1xb4-adamw-40kmtd-512x512.py workdirs/bidnetmscan-t1xb4-adamw-40kmtd-512x512/iter40000.pth

GSD dataset:

python tools/test.py configs/BiDNet/bidnetmscan-t1xb16-adamw-160kgsd-512x512.py workdirs/bidnetmscan-t1xb16-adamw-160kgsd-512x512/iter160000.pth ```

Params and FLOPs

python tools/analysis_tools/get_flops.py configs/BiDNet/bidnet_mscan-t_1xb8-adamw-80k_msd-512x512.py --shape 512

FPS

python tools/analysis_tools/benchmark.py configs/BiDNet/bidnet_mscan-t_1xb8-adamw-80k_msd-512x512.py work_dirs/bidnet_mscan-t_1xb8-adamw-80k_msd-512x512/iter_80000.pth