muwer6 / Method-for-Splitting-the-DeepShip-Dataset

Method for Splitting the DeepShip Dataset

Method for Splitting the DeepShip Dataset

• Welcome to this repository. 😃

• Underwater acoustic target recognition based on deep learning is gradually drawing widespread attention from researchers. However, due to the unique nature of the data, the cost of collection is usually high. Although the use of the DeepShip dataset is becoming increasingly popular, the dataset has not disclosed its method for splitting the training and test sets. This situation makes it difficult for existing methods to be effectively compared. Therefore, we plan to publish the data splitting method used in our paper, in order to provide a reference for subsequent research and to facilitate comparative experiments.

✅ Papers Adopting this Splitting Method

• Paper2: Zhu P, Zhang Y, Huang Y, et al. SFC-Sup: Robust Two-Stage Underwater Acoustic Target Recognition Method Based on Supervised Contrastive Learning[J]. IEEE Transactions on Geoscience and Remote Sensing, 2023.
• Paper1: Zhu P, Zhang Y, Huang Y, et al. Underwater acoustic target recognition based on spectrum component analysis of ship radiated noise[J]. Applied Acoustics, 2023, 211: 109552.

✅ Statement

• We need to declare several details: First, the splitting of the data was done randomly, without any prior knowledge. We hope to conduct comparisons based on this fair approach. Secondly, we endeavor to adhere as closely as possible to the detailed information in the DeepShip dataset paper.

✅ Dataset Splitting Steps

• Download the DeepShip dataset.
• Although the original paper used marine background noise data, it did not disclose this part. Therefore, we have supplemented it with the marine environmental noise data we collected. You can download it through this link (Extraction code:8448).
• The test.txt and train.txt files represent the outcomes of splitting the original recordings. These files reference specific, intact recordings that have not yet been segmented into equally-sized samples.
• The code.ipynb script is responsible for the splitting and segmentation of the dataset. After downloading and organizing the dataset in the your_dir/DeepShip/… format, enter your_dir into audio_path and specify the paths for the two txt files.
• Set the parameters audlen and audstr, where audlen determines the sample length, and audstr specifies the offset between segmentation points. For instance, setting audstr to 32000*3 ensures that the samples do not overlap. Executing code.ipynb produces the segmented results.

✅ Dataset Splitting Results

• When the category of marine environmental noise is used and audstr is set to 32000*1, the ensuing results will be as follows.

	Train	Train	Test	Test	Total	Total
	Sample	Recording	Sample	Recording	Sample	Recording
Cargo	27482	78	10686	31	38168	109
Passenger Ship	31545	120	14303	71	45848	191
Tanker	32330	158	11480	82	43810	240
Tug	26377	42	13965	27	40342	69
No Background	117734	398	50434	211	168168	609
Background	19635	138	8133	59	27768	197
Total	137369	536	58567	270	195936	806

• When audstr is set to 32000*3 and the category of marine environmental noise is excluded, the resulting data will be as follows.

	Train	Train	Test	Test	Total	Total
	Sample	Recording	Sample	Recording	Sample	Recording
Cargo	9185	78	3571	31	12756	109
Passenger Ship	10555	120	4794	71	15349	191
Tanker	10827	158	3857	82	14684	240
Tug	8804	42	4662	27	13466	69
Total	39371	398	16884	211	56255	609

✅ Something Else

• Drawing from our experimental experience, marine environmental noises can be easily distinguished. Hence, we recommend using only four types of ship radiated noise to streamline the training process. Furthermore, setting the sample overlap to zero (audstr=32000*3) is advisable to lessen the computational load.

✅ Ciation

@article{zhu2023sfc,
  title={SFC-Sup: Robust Two-Stage Underwater Acoustic Target Recognition Method Based on Supervised Contrastive Learning},
  author={Zhu, Pengsen and Zhang, Yonggang and Huang, Yulong and Lin, Boqiang and Zhu, Minwen and Zhao, Kunlong and Zhou, Fuheng},
  journal={IEEE Transactions on Geoscience and Remote Sensing},
  year={2023},
  publisher={IEEE}
}

@article{zhu2023underwater,
  title={Underwater acoustic target recognition based on spectrum component analysis of ship radiated noise},
  author={Zhu, Pengsen and Zhang, Yonggang and Huang, Yulong and Zhao, Chengxuan and Zhao, Kunlong and Zhou, Fuheng},
  journal={Applied Acoustics},
  volume={211},
  pages={109552},
  year={2023},
  publisher={Elsevier}
}

✅ At Last

• I hope this dataset splitting method will be helpful to everyone, and I also hope that it will make it more convenient for everyone to conduct comparative experiments 👈. May everyone enjoy smooth sailing and fruitful achievements in their research 😃.