Buškus K et al. (2021), Automated Quantification of Brittle Stars in Se...

ECB-ART-50660

Sensors (Basel) 2021 Nov 16;2122:. doi: 10.3390/s21227598.

Show Gene links Show Anatomy links

Automated Quantification of Brittle Stars in Seabed Imagery Using Computer Vision Techniques.

Buškus K , Vaičiukynas E , Verikas A , Medelytė S , Šiaulys A , Šaškov A .

???displayArticle.abstract???
Underwater video surveys play a significant role in marine benthic research. Usually, surveys are filmed in transects, which are stitched into 2D mosaic maps for further analysis. Due to the massive amount of video data and time-consuming analysis, the need for automatic image segmentation and quantitative evaluation arises. This paper investigates such techniques on annotated mosaic maps containing hundreds of instances of brittle stars. By harnessing a deep convolutional neural network with pre-trained weights and post-processing results with a common blob detection technique, we investigate the effectiveness and potential of such segment-and-count approach by assessing the segmentation and counting success. Discs could be recommended instead of full shape masks for brittle stars due to faster annotation among marker variants tested. Underwater image enhancement techniques could not improve segmentation results noticeably, but some might be useful for augmentation purposes.

???displayArticle.pubmedLink??? 34833671
???displayArticle.pmcLink??? PMC8617961
???displayArticle.link??? Sensors (Basel)
???displayArticle.grants??? [+]

???attribute.lit??? ???displayArticles.show???

References [+] :

Chiang, Underwater image enhancement by wavelength compensation and dehazing. 2012, Pubmed

	Figure 1. Map of the study area where the seabed imagery was collected. Borebukta bay on Spitsbergen Island, Svalbard (Norway).
	Figure 2. Annotation variants considered when preparing seabed mosaics for the semantic segmentation task. (a) Full shape annotation. (b) Disc annotation.
	Figure 3. Example result of underwater image enhancement (left to right, starting from the top row): original raw image, DCP, MIP, RoWS, Paralenz, CLAHE, Fusion, GC, ICM, RGHS, UCM, UDCP, ULAP, TIP2016.
	Figure 4. The input to the network is a patch of the mosaic, the output - semantic segmentation mask. Adapted from Ref. [22]. Since the feature map output by the ResNet-101 backbone is 1/8 the input image size, the pyramid pooling module is followed by upsampling through bilinear interpolation to get a segmentation mask of proper dimensions.
	Figure 5. A generalized overview of sliding window approach to slice an input image into patches for training input.
	Figure 6. Segmentation results when testing on mosaic-1: comparison between acceptable (first two rows) and not so successful (last two rows) mask predictions. (a) Raw image. (b) Ground truth. (c) Prediction from PSPNet.
	Figure 7. Example of false positive prediction which is still useful for counting.
	Figure 8. Example of false negative prediction from the model: organisms with hardly visible disk, marked in green color (left hand side), were missing in the prediction result (right hand side).