Robust Convolutional Neural Network Model For Recognition of Fruits

Deepali M Bongulwar,S N Talbar

doi:10.17485/ijst/v14i45.1493

Abstract

Objectives: To develop a model for the automatic recognition of fruits utilizing deep learning techniques. Methods: We have designed a fruit classification and recognition Model using Convolutional Neural Networks (CNN). We have used excellent quality ImageNet dataset of fruit images for evaluation purpose. It contains 9,130 images of 11 different categories. The classification is challenging as the images comprise different fruits of the same color and shape, overlapped fruits, the background is not homogenous, and with different light effects etc. Findings: We have achieved a validation accuracy of 91.28 % and the testing accuracy of 100%. The same model is trained on the fruits-360 dataset with 92 categories of fruits with 47,526 images. The Model gives validation accuracy of 100% and testing accuracy of 100%. This study also compares the results obtained using transfer learning by training the EfficientNet-b0 architecture with the ImageNet and fruits-360 dataset. The validation accuracy is 96.77% and the testing accuracy is 100%. Again, the validation accuracy and testing accuracy for fruits-360 dataset is 100% and 99.9% respectively. Applications: Recognition of fruit is required in agricultural problems like robot harvesting and fruit counting and many more applications. Moreover, it can be used in the retail business, as a self-service system to recognize the fruits. It can be used in human-robot interactions. Novelty: The model once trained can achieve state-of-art accuracy for the recognition of any type of fruit with any background. It sometimes exceeds human-level performance. Hence, the Model is Robust enough to recognize the fruits. Keywords: Deep learning; Convolutional Neural Networks; Fruit recognition; ImageNet; Transfer learning; Machine learning; fruits-360 dataset; EfficientNet-b0 model

Highlights

The mechanism of automatic classification of fruits gains importance due to labor cost increase and due to the lack of labor in rural production
We proposed a Convolutional Neural Network (CNN) based fruit recognition technique to classify fruit images
ImageNet dataset images are with complicated backgrounds and so could not reach 100% validation accuracy but could find testing accuracy as 100%

Summary

Introduction

The mechanism of automatic classification of fruits gains importance due to labor cost increase and due to the lack of labor in rural production. The overall accuracy obtained was 94.94% over clean dataset but on complicated images like images with varied background the overall accuracy was 89.60%, over unfocused images was 91.03%, and over the occluded images was 92.55%.The authors of(13) proposed a CNN based technique for automatic fruit recognition and classification. They worked on a public dataset and self-made dataset with 7 categories. We proposed a Convolutional Neural Network (CNN) based fruit recognition technique to classify fruit images. Fruits were exposed to light, kept in dark, and some images were black and white

Objectives

Methods

Results

Discussion

Conclusion