Efficient Pre-Harvest Ripeness Estimation Techniques for Fruits

Abstract

The ripeness estimation of fruits plays an important role in marketing and evaluation of quality. However, due to the subjectivity, time consumption and slow speed in case of manual assessment, agriculture industry leads to the need of automation. In this research work, an efficient ANFIS based Pre-harvest Ripeness Estimation (APRE) and Faster R-CNN based Ripeness Estimation (FRRE) techniques have been proposed. In case of APRE, ripeness estimation of fruits is done based on colour. There are three main phases of the APRE: Data Analysis and Processing, Input Feature Selection and Fuzzy Logic Controller deployment. In the first phase, data set of images of fruits is prepared in image acquisition phase. Then images are pre-processed to make them equal in size. In Image Segmentation phase, a fruit is extracted from its background. The two colour features: red-green colour difference and red-green colour ratio are calculated on the basis of the extracted RGB colour attributes. The performance is analysed for these two colour features based on classification accuracy. Adaptive Neuro-Fuzzy Inference System (ANFIS) is utilized for designing and implementing the classification technique which will classify the fruits into four maturity stages. Training dataset is partitioned into four different classes which represent the four different stages of ripeness. In the second proposed technique i.e. FRRE, Faster RCNN has been utilized. Region Proposal Network has been utilized in Faster RCNN to generate the bounding boxes of probable regions where an object of interest can lie. This approach is also much quicker to deploy for new fruits, as it requires bounding box annotation rather than pixel-level annotation. The performances of the proposed techniques are compared with three classification techniques: SVM, Decision Tree and KNN for the classification of the fruits on the basis of the ripeness in terms of specificity, F-Measure, precision, FP Rate, sensitivity and accuracy.