Articles
NEW UNSUPERVISED APPROACH FOR SOLVING CLASSIFICATION PROBLEMS WITH COMPUTER VISION
Article number
562_43
Pages
361 – 375
Language
English
Abstract
A traditional approach at solving classification problems with computer vision is to analyse the objects in question, interview experts and define features that are important for the discrimination between classes.
This approach works very well if the discrimination criteria can be well described by the experts, and captured in the features.
However, this approach has some disadvantages.
Whenever a new problem arises, the procedure has to start over.
Also, when the criteria are beauty and appreciation of pot plants they become very difficult to describe in a consistent way.
A new approach that overcomes the issue of subjectively selecting features for discrimination is suggested in this paper.
It is based on data reduction by means of linear projection into a low dimensional space.
The only decision between raw image and classification of the object is the choice of a linear projection method.
The images of the pot plants are projected onto the eigen-vectors created from training set, and the projected values are used directly for classification.
The eigen-vectors of the training set represents a visualisation of the variance within the pot plants.
The light areas have great importance, hence large variance, while the darker areas have small variance, for instance the background.
This fact calls for an interpretation when this knowledge is coupled with the discrimination information within the specific eigen-vector.
It is, to a point, reversion of the traditional procedure, by exploring the data and theyre by finding the important areas with regard to classification.
This approach works very well if the discrimination criteria can be well described by the experts, and captured in the features.
However, this approach has some disadvantages.
Whenever a new problem arises, the procedure has to start over.
Also, when the criteria are beauty and appreciation of pot plants they become very difficult to describe in a consistent way.
A new approach that overcomes the issue of subjectively selecting features for discrimination is suggested in this paper.
It is based on data reduction by means of linear projection into a low dimensional space.
The only decision between raw image and classification of the object is the choice of a linear projection method.
The images of the pot plants are projected onto the eigen-vectors created from training set, and the projected values are used directly for classification.
The eigen-vectors of the training set represents a visualisation of the variance within the pot plants.
The light areas have great importance, hence large variance, while the darker areas have small variance, for instance the background.
This fact calls for an interpretation when this knowledge is coupled with the discrimination information within the specific eigen-vector.
It is, to a point, reversion of the traditional procedure, by exploring the data and theyre by finding the important areas with regard to classification.
Authors
L. Kohsel
Keywords
Online Articles (50)