Date

2018-07-05

Author

Bircanoglu, Cenk
Atay, Meltem
Beser, Fuat
Genc, Ozgun
Kizrak, Merve Ayyuce

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

382
views

0
downloads

Cite This

Waste management and recycling is the fundamental part a sustainable economy. For more efficient and safe recycling, it is necessary to use intelligent systems instead of employing humans as workers in the dump-yards. This is one of the early works demonstrating the efficiency of latest intelligent approaches. In order to provide the most efficient approach, we experimented on well-known deep convolutional neural network architectures. For training without any pre-trained weights, Inception-Resnet, Inception-v4 outperformed all others with 90% test accuracy. For transfer learning and fine-tuning of weight parameters using ImageNet, DenseNet121 gave the best result with 95% test accuracy. One disadvantage of these networks, however, is that they are slightly slower in prediction time. To enhance the prediction performance of the models we altered the connection patterns of the skip connections inside dense blocks. Our model RecycleNet is carefully optimized deep convolutional neural network architecture for classification of selected recyclable object classes. This novel model reduced the number of parameters in a 121 layered network from 7 million to about 3 million.

Subject Keywords

Recycling , Training , Convolutional Neural Networks , Machine Learning , Computer Architecture , Biological System Modeling , Metals

URI

https://hdl.handle.net/11511/67997

Collections

Graduate School of Natural and Applied Sciences, Conference / Seminar