Transfer Learning based Fuzzy Deep Neural Networks for leaves detection from Digital Images

Abstract

Automatic detection of leaves from digital images has become important technique for identifying phenotypic changes in plants. Application of Machine learning concepts for automatic detection of leaves from images is the latest advancement in computer vision. Deep neural networks (DNNs) such as Google Nets, Alex Nets and Mobile Nets which belong to machine learning concepts are known for identifying the leaves in an image. The limitation of existing DNNs is that they do not handle uncertainty in the image during the classification stage. Class Wise Belongingness granulation of input image would effectively handles the uncertainty and improves the accuracy of classifier. In the present study, we propose a Transfer learning based Fuzzy Deep Neural Networks (TLFDNNs) model for identifying the leaves in digital Images. In the proposed model, the input image is fuzzy granulated based on class wise belongingness (CWB). Furthermore, the leaves in fuzzy granulated image are detected using Mobile Nets. The CWB based granulation of proposed model produces better results in comparison with conventional deep neural network models such as Google Nets, Alex Nets and Mobile Nets. The improvement in performance of TLFDNN model over other type of deep neural network models is justified by testing on three leaf image datasets such as Citrus, Azadirachta indica and Psidium guajava. The performance of models was evaluated using the metrics like average percentage of leaves detected in an image and the standard deviation of average percentage of leaves detected in the test images.