Fused random pooling in convolutional neural network for herbal plants image classification

Abstract

Convolutional Neural Network (CNN) faces concerns on overfitting. The CNN model learns during the training process but may not be able to classify new data correctly. Hence, the accuracy is higher in the training set than in the validation set. This occurs despite the breakthrough in CNN even if it is considered state of the art in image analysis. In this study, the fused random pooling is presented to create enhanced pooled feature maps, in that way, reducing overfitting and improving classification accuracy. The dataset utilized in this study consists of 8686 of six herbal plant images collected by actual photos and gathered publicly available online images. A comparison in terms of validation accuracy and validation loss of the average, max, mixed, and fused random pooling methods is presented. Results show that the fused random pooling achieved the highest validation accuracy of 98.21 percent and the lowest validation loss of 5.57 percent among the pooling methods used. The fused random pooling also led in terms of the performance evaluation on the precision, recall, and F1 score and attained results of 98.23 percent, 98.21 percent, and 98.21 percent, respectively. Thus, proving that applying fused random pooling is reliable in the herbal plant image classification.