en

Development

Detection of important areas in the image

In cooperation with the Faculty of Mathematics, Physics and Informatics at the Comenius University, we conducted a research project on the detection of important areas in the image, resulting in software useful in practice in many areas of life, such as image compression, quality evaluation of image and video, object detection, retargeting, creating thumbnails and many more.

The research is based on the fact that the human visual perception and processing of huge amounts of visual information (from 108 to 109 bits per second) is a very difficult task for a person. One cannot see everything around them, but some things attract visual attention reflexively (independent of a role), other things intentionally, depending on certain tasks (selective attention).

Visual attention

 Key:
1 – Free independent view
2 – Looking at the task to establish the age of the figures in the image

3 - Looking at the task to establish the activity they did before the arrival of an unexpected visit

4 - Looking at the task to memorise what the figures were dressed in

5 - Looking at the task to memorise the position of the figures and objects in the room

6 - Looking at the task to estimate how long the visitor(s) will stay

 

The goal of the research was therefore to reduce as effectively as possible the amount of this information by means of visual attention, which is used for selection of areas in the scene (the important areas), and also detect the main areas of sensations by means of a PC.

For determining significant areas we used computational models, which are normally based on the detection of symptoms in the image and the subsequent combinations, while the final map of important areas represents a grey level map.

In the original model, which is based on a combination of symptoms (colour, intensity and texture), the authors have proposed a new strategy regarding the combination of these symptoms, which suppresses the incorrectly marked areas as important in a contrasting map. This strategy uses the local contextual information to suppress the disputed areas and highlights the true distinctive zones. The system or model used by us is a superset of the original model and lies in the detection of 4 symptoms (colour, intensity, texture and key points - DLR). These symptoms are then combined into a single map and the resulting map of important areas is created through the suppression of wrongly

Comparison of results of individual methods

a) original image

b) map of important areas using the DLR function

c) Itti - model

d) CSI - model

 

 

To Top