Recycling waste is necessary for the sustainability of a good social life. The recycling process for waste is generally carried out manually and with many types of waste categories. For the waste sorting process to run more effectively and efficiently, automatic waste sorting is required. One example is by using an image-based waste type classification system using deep learning or machine learning algorithms. In previous research, AlexNet and SVM were used as classification algorithms for waste types in 6 different waste categories. The results of this study show that SVM performance is better with an accuracy of 63% compared to AlexNet's performance with an accuracy of 20%. Whereas in general, the CNN algorithm's performance should produce better accuracy than SVM. Based on this, we propose to re-examine the classification of waste types in the same dataset as previous research using the latest CNN algorithm architecture, namely ResNet. More completely, the architectures used are AlexNet, ResNet-18, and ResNet-50, respectively with and without pretrained, so that a total of 6 types of CNN algorithm architecture are used in the training. In this study, the AlexNet architecture is used with a different configuration from previous studies. The test parameter is the number of epochs of 50 epochs with attention to how many epochs the training results have shown convergent. From the training results, the highest accuracy obtained by the ResNet-50 architecture with pretrained is 91.16% and shows convergent results since the 14th epoch. Then for the lowest accuracy obtained by the AlexNet architecture without pretrained, which is 58% and shows convergent results since the 21st epoch.
Copyrights © 2022