Dzalfa Tsalsabila Rhamadiyanti, Kusrini

Analisa Performa Convolutional Neural Network dalam Klasifikasi Citra Apel dengan Data Augmentasi

Introduction

Analisa performa convolutional neural network dalam klasifikasi citra apel dengan data augmentasi . Analisis performa Convolutional Neural Network (CNN) untuk klasifikasi citra apel dengan augmentasi data. Temukan teknik noise injection capai akurasi 98.82% dan dampaknya.

Abstract

Augmentation is creating new samples from an original dataset by applying small random transformations to the original dataset but retaining its labels. This research applies Data Augmentation to the Convolutional Neural Network model for apple image classification. The apple images used are Braeburn apples which have orange to red skin with a yellow background, Crimson Snow apples which have red skin, and Pink Lady apples with bright pink skin and yellow and green hues. There are 675 apple images used, divided into three classes, each with 225 photos. Four augmentation techniques are applied, namely flipping, cropping, rotation, and noise injection. This research carried out six scenarios, namely without augmentation, using each augmentation technique separately and combining two augmentation techniques, which produced the highest accuracy values. From the six scenarios, it was found that the augmentation technique that produced the best accuracy value was noise injection, namely 98.82%, followed by flipping with an accuracy of 72.78%, then rotation with an accuracy value of 68.64% and an augmentation technique that produced an accuracy value. The lowest is cropping, namely 67.46%. The two best augmentation techniques, noise injection, and flipping, were combined and produced an accuracy value of 84.02%. The accuracy value obtained by this combination could be more optimal due to the effect of noise injection, which can erase consistent changes in orientation from flipping. This needs to be improved so that the model can learn consistent features. It is hoped that future research can maximize the effectiveness of augmentation techniques by choosing augmentation techniques that complement each other and suit the characteristics of the data being processed

Review

This paper investigates the efficacy of various data augmentation techniques for improving Convolutional Neural Network (CNN) performance in classifying apple images. Focusing on three distinct apple varieties – Braeburn, Crimson Snow, and Pink Lady – the research applies flipping, cropping, rotation, and noise injection, individually and in combination, to a dataset of 675 images. The study's objective to optimize CNN performance through data augmentation is highly relevant in agricultural applications and for addressing data scarcity challenges inherent in many real-world machine learning problems. The clear definition of the dataset and the specific augmentation techniques chosen provides a solid foundation for evaluating their impact. The methodology explores six scenarios, culminating in a comparison of individual augmentation techniques and a combination of the two best performers. A striking finding is the exceptionally high accuracy of 98.82% achieved solely with noise injection, which significantly outperforms other individual techniques (flipping 72.78%, rotation 68.64%, cropping 67.46%). This result is particularly noteworthy as noise injection often serves as a regularizer rather than a primary performance booster to such an extent, warranting further discussion on the specific characteristics of the noise applied and its interaction with the CNN model and the apple image features. Moreover, the subsequent drop in accuracy to 84.02% when combining noise injection with flipping is a critical observation, suggesting that not all effective individual augmentations are additive, with the authors' explanation regarding noise potentially "erasing consistent changes" offering a plausible hypothesis for this interaction. While the research successfully identifies noise injection as a highly effective standalone augmentation for this specific task, the abstract highlights a key challenge in optimally combining techniques. Future work should delve deeper into the nature of the noise applied and how it contributes to such high accuracy, potentially exploring its regularization properties or how it enhances specific feature learning. A more detailed analysis of why other standard techniques performed relatively poorly, and why the combination led to a decrease in performance compared to the best individual technique, would be highly beneficial. This could involve examining the feature space implications of each augmentation and developing strategies for selecting truly complementary techniques, as suggested by the authors, to ensure the model learns consistent and robust features rather than being overwhelmed by conflicting transformations.

Full Text

Comments

You need to be logged in to post a comment.