Article Preview
Top1. Introduction
Tomato is considered most common crop, consumed at large scale in India. It is a climacteric vegetable that continues to ripen even after it has been harvested. It is possible to determine the maturity level of the tomato by simply observing its surface color (Choi et al., 1995). However, the manual classification of tomatoes is very cumbersome, error prone and tiring due to involvement of human factor. In addition to this, weaken interest of young people in agriculture has led to hike in the production cost of many agricultural products in India. Owning to this, prices of tomato rises significantly during peak season. In current scenario, the deep learning based tomato maturity assessment methods could be of immense help to address the above stated problems (Li et al., 2010) (Guo et al., 2015).
Tomato possesses several attributes such as size, colour, and shape which can be utilized to classify it after gaining maturity. However, due to visual appearance being an important aspect of quality assessment, colour is considered to be most preferred attribute for determining the tomato’s maturity level. The surface colour reflects the quality of tomatoes influencing the customer’s preference. Hence, colour feature might play an important role in developing an automated maturity grading system for classification of tomatoes.
These days, tomato classification is done manually by skilled labourers based on their physical quality. However, manual classification is very expensive and their assessment easily gets affected due to shortages of manpower during peak seasons. Moreover, it is error prone and time-consuming process due to involvement of human being. A global marketing strategy as outlined by the World Trade Organization (WTO) expects high-quality marked food items. Therefore, currently development of an efficient maturity grading system is in high demand to lower down the production cost and improve the quality of the assessed items (Arah et al., 2016; El-Ramady et al., 2014; Londhe et al., 2013; Pavithra et al., 2015; Syahrir et al., 2009).
Over the past twenty years, the deep learning methods have gained wide popularity for image classification (Alam et al., 2021), computer vision, and pattern recognition tasks (Singh et al. 2021). The deep learning methods have also gained wide acceptance in the agriculture field due to better classification capability of these methods as compared to traditional image processing and machine learning methods (Abiodun et al., 2018; Alom et al., 2019). Image processing and machine learning methods are dependent upon manual extraction and fine tuning of features which makes conventional methods labour intensive and time consuming. Whereas deep learning methods depend upon the automatic extraction of features which derive valuable information from a large volume of data (Kamilaris & Prenafeta-Boldú, 2018).
Major concern with deep neural network (DNN) model is over-fitting problem, however its remarkable efficiency in many vision related task makes CNN suitable for classification of agricultural products especially for tomato’s classification and fruit classification problems. In CNN, over-fitting problem mainly occurs due to three factors, i.e. extremely complicated architectures, data interference, and insufficient training set. Data augmentation is one of the possible solution to minimize the over fitting problem in case where data is small in quantity (Abiodun et al., 2018). In addition, the strong hierarchical organizational structure as well as the huge potential of the CNN models makes them more competent to solve complex problems.