A deep CNN approach to detect and classify local fruits through a web interface

Abstract

An automated fruit identification system may be used in the supermarket to help consumers determine the type and standard of fruits. Automatic fruit categorization and recognition is becoming increasingly popular, but it is also becoming more difficult due to low contrast and ambiguous characteristics. Proposed research used deep learning algorithms to classify fruit images and developed a web-based system for autonomous fruit identification. However, autonomous fruit categorization is not a relatively easy process that is dependent on the locations, shapes, colors, and sizes of the objects. Proposed study collected samples from different local places, then removed images backdrops and improved them for a more accurate result. ResNet-50, VGG-19, Inception-V3, and MobileNet were utilized to achieve more precise feature extraction. Among these, MobileNet achieved 99.21% accuracy in feature extraction, outperforming previously proposed machine learning techniques. 3240 samples of eight different fruits are collected from the Bangladeshi countryside, including Carambola, Bilimbi, Elephant Apple, Emblica, Burmese Grape, Sapodilla, Tamarind, and Wood Apple. As a consequence of this research, the proposed approach will aid individuals and industry in recognizing local fruits due to its high accuracy rate and web-based system.

Keywords

• Deep CNN
• Pre-trained models
• Fruit classification
• Web interface