Improvement of Multichannel Image Classification by Combining Elementary Classifiers
Владимир Васильевич Лукин, Галина Проскура, Irina Vasilyeva · 2019
A post-classification processing technique for multi-channel images that includes three stages is proposed. The purpose of the first stage is to correct decisions of pixel-by-pixel classifiers based on estimates of classes' posterior probabilities. At the second stage, a logical convolution of the classification layers is performed which makes it possible to select the most probable class. At the final stage, local spatial filtering of pre-segmented image is done which is performed in the neighborhood of detected segments' edges. The post-classification processing effectiveness is verified for satellite images. It is demonstrated that the proposed post-classification processing procedure can significantly increase the probability of recognizing poorly distinguishable classes and improve overall accuracy of image classification.