Home

Awesome

resim

VHRShips: An extensive benchmark dataset for scalable deep learning-based ship detection applications

This study focuses on all stages of ship classification in optical satellite images. We proposed a deep learning-based “Hierarchical Design (HieD)” approach performing Detection, Localization, Recognition, and Identification (DLRI) of the ships in optical satellite images. HieD is an end-to-end approach that allows the optimization of each stage of the DLRI independently. We generated a unique and rich ship dataset (Very High Resolution Ships, VHRShips) using Google Earth images including diversity of ship types and data acquisition conditions. While the Xception network is used in detection, recognition, and identification stages; YOLOv4 is preferred for the localization of the ships. We found F1-Scores for DLRI stages are respectively 99.17 / 94.20 / 84.08 / 82.13 percent; end to end F1-Scores are 99.17 / 93.43 / 74.00 / 57.05 for the same order.

The Steps to be Followed to Run the Code:

  1. Download "matlabCodes" and "YOLOV4" folders
  2. Download "testList.mat", "trainedNets_Weights" and "test_c34.7z" (test images) from https://drive.google.com/drive/folders/1cl0vQa78Yhslxdry5QGHUOhqNCYPQXjx?usp=sharing
  3. Open MATLAB
  4. Add the path of downloaded files at the step 1 and 2
  5. open "sK_HieD_fileRootFixerV1.m", execute the orders defined with ">>>" sign and, execute the script
  6. open "sK_HieD_mainV1.m", execute the orders defined with ">>>" sign and, execute the script
  7. open "sK_HieD_loopV1.m", execute the orders defined with ">>>" sign and, execute the script
  8. open "sK_HieD_evaluatorV1.m", execute the orders defined with ">>>" sign and, execute the script
  9. execute "sK_HieD_visualizeV1.m"

Details about testList.mat

File is a table in 1275*37 (rows and columns) format. The first column is the image address. The second column is the list of all bounding boxes in the related image. The third column is the list of ship labels in the related image. The columns between 4th and 37th involve the bounding boxes of the the related ship class if exists in the related image. resim

Dataset

VHRShips comprises of 6312 optical satellite images collected from Google Earth. While 1000 images don’t include any ship, rest of them consist 11337 ships in 34 different classes. VHRShips has high definition, inshore and offshore images from various locations, rich metafile content and includes noisy images representing real scenarios.

VHRShips metafile includes the following columns with the described contents
Image Order : 1-6312
Image Location : (list of the locations given in the article)
Image Date : YYYYMMDD (Year Month Day)
Image Quality : 1-3 (3: good (clear); 2: moderate (low resolution, glint); 1: poor (cloudy, haze))
Ship Existence : 0 / 1 (0: non-exist; 1: exist)
Number of ship(s) : 0-118
Length of the longest ship: 7-408 meters
Explanation_1 : 1-7 (1: Alongside vessel, 2: Ship in ballast, 3: Open hatch, 4: Be towed, 5: Discharging water, 6: Tagging, 7: At pier)
Explanation_2 : Free style comment
Bounding Box : x, y, width, height (bounding box information of ship(s) distributed to the classes)

RELATED PUBLICATION: Kızılkaya S, Alganci U, Sertel E. VHRShips: An Extensive Benchmark Dataset for Scalable Deep Learning-Based Ship Detection Applications. ISPRS International Journal of Geo-Information. 2022; 11(8):445. https://doi.org/10.3390/ijgi11080445

NOW full dataset is open access. Please cite the journal given above if you use the dataset. Use the link below to download VHRShips, annotations and meta file. https://drive.google.com/drive/folders/16IBRXXgzL2l-Xvaaxxu9NUnHPRoML6d8?usp=sharing

The metafile will be updated soon.