Impact of Activation Functions and Number of Layers on the Classification of Fruits using CNN

Abstract

The classification of fruits into various classes is becoming inherent to the food processing industry. This paper presents a simulation analysis of effect of activation functions: ReLu, Softmax, sigmoid, and Softplus on the accuracy and latency of the CNN algorithm for classification of fruits: apple, banana and orange. The paper presents the comparative increase of accuracy of different activation functions over the ReLu activation function. The algorithm is trained and tested over a database created by downloading fruit images from the online sources. Also, this paper presents the effect of increasing the number of convolutional layers of the CNN algorithm on the Accuracy and latency of the model. The software used for simulation of the model is Python implemented using Jupyter Notebook over the Anaconda platform.