BioSpecular Dataset: Benchmark for specular reflection removal in endoscope image

We present a dataset of images of dissected lung tissue, with and without (with reduced) specular reflection, to validate specular reflection removal techniques for endoscope images.

This dataset is part of our paper titled: "BioSpecular: A Benchmark of Specular Reflection Removal on Tissue Images" which will be published in 6th IEEE International Conference on ICT Solutions for eHealth.

To create this dataset, a pair of specially designed endoscope-like probes were used to images a dissected lung from an ovine sample. One of the probe capture image of the tissue with specular reflection, meanwhile the other probe which has polarisers in orthogonal configuration capture the image of the tissue with reduced specular reflection. The experiment was designed such that images from both of the probes were co-located. Furthermore, existing streo-correspondence algorithms were used to enhance the co-location of the images from both probes. Hence, the image pairs, with normal and reduced specular reflections, can be used to validate specular reflection removal techniques. In our particular publication, we compares several state-of-the-art deep learning-based video specular reflection removal techniques. Details on the comparison study, as well as the procedure of the experiment to create the dataset can be found in the publication.

This dataset comprised of two sub datasets:

• Pairs_Dataset.zip

This sub dataset consists of the following folders:

crosspol: Images of lung tissue (N=15) with reduced specular reflection acquired using the cross-polarisation probe, 

crosspol_final: Images from crosspol with enhanced co-location.

full_videos: Videos of lung tissue with normal amounts of  specular reflection acquired using the probe without the cross-polarisation. A frame from either the beginning and/or the end of each video is co-located with its corresponding image in (1). The length of each video is varied between 200-1000 frames. 

for_video_inpaint: The cut version of videos in full_videos which has only the first or last 150 frames containing the co-located frame (N=15). The cut version was used for comparison study of several specular reflection removal techniques, as detailed in our publication. We divide the videos into to two: the finetuning videos which were used for finetuning and validating a specular reflection detection model; and the test videos which were used to test the detection and inpainting model.

glass: Frames from each video in for_video_inpaint which were co-located with images from crosspol_final. 

weak_annotation: Detection mask for specular reflection from images in glass acquired using K-means clustering and histogram thresholding. Parameters used for the annotation techniques can be found in csv files within. The code that was used to create the weak annotation was published in our github repository which is listed in this page.

The name format for each image in glass, crosspol_final, and weak_annotation is posxx_yyyyy.png, where xx is the index for sample area and yyyyy is the number of the frame within the video posxx in for_video_inpaint and full_videos.

• Keyframes_Dataset.zip

For our comparison study, we used a deep learning-based detection model to first detect the specular reflections in the video before removal. This sub dataset contain keyframes extracted from videos pos01-pos04 in the finetuning folder in for_video_inpaint as well as additional keyframes from videos of ovine muscle samples. The keyframes were weakly annotated for specular reflection using K-means clustering and hitogram thresholding.