An integrated framework for processing of fruits and vegetables using image analysis and machine learning techniques: a De Novo approach

Abstract

Abstract Fruits and vegetables are necessary for our daily life as it contains many important nutrition constrains. Hence, fruits and vegetables are the most profitable agricultural product. The appropriate selection of suppliers is very important for a retail store of fruits and vegetables. The fruits and vegetable processing includes lots of tasks from harvesting to reach the customer's hand. The tasks are segmentation, sorting, classification, grading, etc. The manual execution of those tasks is very time-consuming and requires a huge number of expert resources. Hence, this research aims to propose a framework for automating those tasks using image analysis and machine learning techniques. This work proposes an end-to-end framework for fruits and vegetable processing in agricultural industries as well as in the supermarket. The fruits and vegetable segmentation is mandatory for further processing. Segmentation and detection are also important for harvesting fruits and vegetables from the natural environment. A graph-based segmentation technique was used to segment the foreground fruit or vegetable object from the natural background. A region of interest detection technique is also proposed here. The experimentation shows that the proposed technique performs better than the popular Otsu thresholding technique in this context. The non-fresh i.e. rotten or defective fruits and vegetables are very harmful to fruits and vegetables inventory or store. The non-fresh fruits and vegetables need to be detected and removed as early as possible. The computer vision system has to rely only on the visual features to label fruit and vegetable as rotten or defective. Another challenge is that the pattern of rot and defect is different for different types of fruits or vegetables. A convolutional neural network architecture has been proposed for the classification of fresh and non-fresh fruit and vegetables. The test accuracy and f1 score of the proposed network architecture are 97.74% and 98.43% respectively. There is an almost infinite number of fruits and vegetable species in the world. Hence, the classification of fruits and vegetables using visual features is a very difficult task. A novel framework has been proposed here for the classification of fruits and vegetables using image analysis and machine learning techniques. The shape, color, and texture features are combined to classify fruits and vegetables. An exploratory analysis has been done on the performance of different classification algorithms in this context. The experimentations are done on 35 classes of edible fruits and vegetables. The proposed framework achieves at most 99.96% classification accuracy. The grading is very necessary to get the proper market value of fruits and vegetables. There are different grading parameters i.e. shape, size, color, volume, mass, etc. Volume and mass estimation is more challenging using computer vision than the estimation of other parameters. The challenge increases when the shape of the fruit or vegetable is irregular or non-axisymmetric. Here, a novel split and merge technique has been proposed to estimate the mass and volume of fruit and vegetable from a single image. The experimentation has been done on both regular and axisymmetric as well as irregular and non-axisymmetric fruit and vegetable. The results have been validated with water displacement and digital balance for volume and mass respectively. The overall correlation for volume and mass estimations are 0.96 and 0.97 respectively. The selection of an appropriate supplier is very important to reduce the complexity in further processing of fruits and vegetables. A holistic framework for fruit and vegetable supplier selection is proposed in this thesis. The proposed framework integrates principal component analysis and technique for order of preference by similarity to ideal solution. The proposed approach reduces the number of criteria as well as correlations among the criteria. It also generates weights through the principal component analysis for each of the newly generated criteria. The framework finally ranks among the suppliers.