Meta-learning for Image Processing Based on Case-Based Reasoning - Computational Intelligence in Healthcare: Advanced Methodologies - page 236

Even the automatic evaluation has some weak points. The automatic evaluation
proposed in [10] takes the original images, produces an edge image out of it based on
the chosen standard edge operation, and labels the edges based on a simple threshold-
ing algorithm (see Fig. 23d, with Otsu´s thresholding algorithm) and then compares
this image with the output image of the Watershed Transformation based on case-
based reasoning according to Zamperoni´s similarity measure [18]. For each of these
processing steps different standard image processing operations can be used, produc-
ing different results for the same image.

(0.75, 2, 1) BS=96 (1.25, 0.75, 1) BS=92 (0.5, 0.75, 1) BS= 73

(a) Image parrot (b) Manually evaluated

by a human_1

(c)Manually evaluated
by a human_2

(d)Automatically
evaluated segmentation
result

Fig. 23. Different manually evaluated and automatically evaluated segmentation results of
image parrot

5 Elicitation of Image Descriptions and Assessment of Similarity
for Watershed Transform

The aim of the image description is to find out from a set of images the group of im-
ages that needs the same processing parameters for achieving the best segmentation
results. To give an example, the images neu1, neu3, neu4 and parrot should be
grouped together based on the best parameters a, b, and T.
We consider different image descriptions in our study that should allow us to group
images based on the image parameters and by doing so learn a model for image seg-
mentation from samples.
Cases are normally composed of

• non-image information

• parameters specifying image characteristics, and

• parameters for solution (image segmentation parameters).

Non-image information is different depending on the application. In our study we
use images from different domains like landscape, faces or biology. Our aim is to
describe image similarity only by general image properties. Hence our cases are com-
posed of

• parameters specifying image characteristics and

• parameters a, b, T for segmentation.