Comparison of Increase in Accuracy of Classification in Image Recognition Models Using Different Gaussian Filtering Algorithms Under Gaussian Noise

Abstract

Many studies focus on comparing two different noise filtering algorithms, to observe their effectiveness under different noise conditions However, not a lot of studies consider the varying effectiveness between different variants of the same type of algorithm. The aim of the study is to show how effective each Gaussian different filter variant is at reducing Gaussian noise by assessing the accuracy, and moreover finding which Gaussian filter variant increases how sure the model is with its classification by assessing the confidence level. A Random Forest Classifier was trained using a controlled image test data dataset. Each image had Gaussian noise applied to them and then was processed through different Gaussian filter variants before getting classified as test data by the model. The results show that Gaussian filters improve the performance of the model by reducing Gaussian noise — increasing the accuracy from approximately 46% to an average value of approximately 50%, when anomalies are removed. It also increases the confidence level of each classification from an average of 33% to an average value of approximately 45%. Furthermore, it was concluded from the results that, the most effective Gaussian filter variant at increasing the accuracy and the confidence level of the classifications was the were two different filters: the Gaussian pyramid de-noise for accuracy with approximately 54%, and the recursive Gaussian filter for confidence with an average of 56%. Overall, the best performance across both fields was the Gaussian pyramid de-noise filter variant with a 54% for accuracy (highest accuracy recorded) and a 50% in confidence.