Authors:
Md Rahman
1
;
JungHwan Oh
1
;
Wallapak Tavanapong
2
and
Piet C. de Groen
3
Affiliations:
1
Dept. of Comp. Sci. and Eng., University of North Texas, Denton, TX 76203, U.S.A.
;
2
Computer Science Department, Iowa State University, Ames, IA 50011, U.S.A.
;
3
Division of Gastroenterology Hepatology and Nutrition, University of Minnesota, MN 55455, U.S.A.
Keyword(s):
CBIR, Deep Learning, Depth Map, Colonoscopy, Vision Transformer (ViT).
Abstract:
Content Based Image Retrieval (CBIR) finds similar images given a query image. Effective CBIR has been actively studied over several decades. For measuring image similarity, low-level visual features (i.e., color, shape, texture, and spatial layout), combination of low-level features, or Convolutional Neural Network (CNN) are typically used. However, a similarity measure based on these features is not effective for some type of images, for example, colonoscopy images captured from colonoscopy procedures. This is because the low-level visual features of these images are mostly very similar. We propose a new method to compare these images and find their similarity in terms of their surface topology. First, we generate a grey-scale depth map image for each image, then extract four straight lines from it. Each point in the four lines has a grey-scale value (depth) in its depth map. The similarity of the two images is measured by comparing the depth values on the four corresponding lines
from the two images. We propose a function to compute a difference (distance) between two sets of four lines using Mean Absolute Error. The experiments based on synthetic and real colonoscopy images show that the proposed method is promising.
(More)