I.J. Intelligent Systems and Applications, 2021, 2, 52-61
Published Online April 2021 in MECS (http://www.mecs-press.org/)
DOI: 10.5815/ijisa.2021.02.04
Copyright © 2021 MECS I.J. Intelligent Systems and Applications, 2021, 2, 52-61
Implementation of Transfer Learning Using
VGG16 on Fruit Ripeness Detection
Jasman Pardede
Department of Informatics Engineering, Institut Teknologi Nasional Bandung, Bandung, Indonesia
E-mail: jasman@itenas.ac.id
Benhard Sitohang, Saiful Akbar, and Masayu Leylia Khodra
School of Electrical Engineering and Informatics, Institut Teknologi Bandung (ITB), Bandung, Indonesia
E-mail: {benhard, saiful, masayu}@informatika.org
Received: 12 June 2020; Accepted: 06 December 2020; Published: 08 April 2021
Abstract: In previous studies, researchers have determined the classification of fruit ripeness using the feature
descriptor using color features (RGB, GSL, HSV, and L * a * b *). However, the performance from the experimental
results obtained still yields results that are less than the maximum, viz the maximal accuracy is only 76%. Today,
transfer learning techniques have been applied successfully in many real-world applications. For this reason, researchers
propose transfer learning techniques using the VGG16 model. The proposed architecture uses VGG16 without the top
layer. The top layer of the VGG16 replaced by adding a Multilayer Perceptron (MLP) block. The MLP block contains
Flatten layer, a Dense layer, and Regularizes. The output of the MLP block uses the softmax activation function. There
are three Regularizes that considered in the MLP block namely Dropout, Batch Normalization, and Regularizes kernels.
The Regularizes selected are intended to reduce overfitting. The proposed architecture conducted on a fruit ripeness
dataset that was created by researchers. Based on the experimental results found that the performance of the proposed
architecture has better performance. Determination of the type of Regularizes is very influential on system performance.
The best performance obtained on the MLP block that has Dropout 0.5 with increased accuracy reaching 18.42%. The
Batch Normalization and the Regularizes kernels performance increased the accuracy amount of 10.52% and 2.63%,
respectively. This study shows that the performance of deep learning using transfer learning always gets better
performance than using machine learning with traditional feature extraction to determines fruit ripeness detection. This
study gives also declaring that Dropout is the best technique to reduce overfitting in transfer learning.
Index Terms: Fruit ripeness, transfer learning, MLP, overfitting, accuracy.
1. Introduction
Indonesia, as an agricultural country, has several agricultural products included mango, apple, orange, banana,
durian, tomato, and others. Good quality fruit will have good selling points [1]. It is necessary to maintain the quality of
fruit consistently to increase the selling value of the fruit. The fruit quality determined by the level of fruit maturity [2],
so we need a system that can be determining fruit quality well. Determination of fruit quality by using the human senses
often does not produce consistent quality and requires a long time. Besides that, to delegate one's expertise to others in
determining the maturity of the fruit requires a long learning process. Therefore, we need tools that can help determine
effectively and efficiently fruit maturity.
Today, fruit recognition systems have become a challenging topic in the field of computer vision [3,4]. Various
image analysis techniques have been developed to help and facilitate work, including automatic fruit harvest detection
systems [5], automatic fruit recognition [6,7], fruit classification system [8], ripeness fruit system [9], fruit diseases
detection system [10], and others. Generally, fruit recognition systems use the feature descriptor, namely the shape
features [11], texture features [12], and color features [2,11-13]. The feature descriptor algorithm used to extract mainly
information from the image stored in a feature vector [3,4,9]. But the performance of the feature descriptor, in the fruit
recognition system, results in suboptimal performance [13]. To improve image classification performance requires an
effective/robust feature descriptor.
One of the best image feature descriptors in the field of machine learning and computer vision that has had great
success to date is convolutional neural networks. Several researchers have utilized CNN as feature extraction in various
applications. Feature extraction with CNN has proven to be very robust. CNN method has a modification in the form of
Deep Convolutional Neural Network (Deep CNN) that is the beginning of deep learning. Digital image processing
methods have been applied in the deep learning classification method also. Deep Learning has become one of the hot