A CNN-based implementation of fruit recognition

Abstract

Image recognition technology is now widely used in various industries, and CNN has played an indispensable role in it over the past decade. The paper discusses the use of Convolutional Neural Networks (CNN) for fruit image recognition, aiming to verify the impact of different CNN designs on model training time, test accuracy, and test accuracy. The experiment uses data from the Kaggle Fruits 360 project and focuses on ten different categories of fruit. The CNN model is built using 3*3 convolutional kernels and features four combinations of convolutional and relu layers. The final test accuracy is recorded as 98.1714%. The paper also discusses potential reasons for lower-than-expected accuracy and attempts to address these issues, including overfitting, image resolution, and the simplicity of the training set. The impact of regularization and different image resolutions on model accuracy is observed. The paper concludes by highlighting the practicality of CNN in image recognition, but also acknowledges limitations such as training time, computational resources, and the abstract nature of extracted features. It also emphasizes the importance of choosing an appropriate training set for model accuracy and suggests that other AI models may offer solutions to the shortcomings of CNN. Overall, the paper provides insights and experiences for those working with CNN in image recognition and acknowledges the rapid development of artificial intelligence in recent years.