Abstract
Image classification is one of the fundamental steps in digital image processing. Research in this area has received considerable attention, with photos shared on social media which are sometimes similar but have different identities. There are various classification methods proposed in the literature to improve accuracy. One important strategy is Convolutional Neural Networks (CNN). Although CNN is superior in pattern recognition, it has limitations inaccuracy. It requires additional training time, especially when dealing with variants in data generated from a large number of images but of similar properties. Therefore, this study aims to overcome this problem by proposing a modification of the CNN layer to increase the accuracy of the multi-class image classification. This research used four different flower species with similar patterns added from a public database. Each category consists of 400 colour images with different angles, backgrounds, and lighting conditions that provide different variations to the training process. Through experiments using 1,600 of the four flower species, this study shows that the 18-34 layer modification produces the most optimal accuracy in the training process ranging from 99.3% with misclassification of MSE =0.0025, RMSE = 0.1606, and MAE = 0. 0133. Meanwhile, the computation time required to compile the data set is 3 minutes, 18 seconds. This result is 50% faster when compared to computation time using existing architecture such as Alexnet model with a similar number of layers.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have