3D Convolutional Neural Networks Initialized from Pretrained 2D Convolutional Neural Networks for Classification of Industrial Parts
Deep learning methods have been successfully applied to image processing, mainly using 2D vision sensors. Recently, the rise of depth cameras and other similar 3D sensors has opened the field for new perception techniques. Nevertheless, 3D convolutional neural networks perform slightly worse than ot...
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2021-02-01
|
| Series: | Sensors |
| Subjects: | |
| Online Access: | https://www.mdpi.com/1424-8220/21/4/1078 |
| _version_ | 1797414777960529920 |
|---|---|
| author | Ibon Merino Jon Azpiazu Anthony Remazeilles Basilio Sierra |
| author_facet | Ibon Merino Jon Azpiazu Anthony Remazeilles Basilio Sierra |
| author_sort | Ibon Merino |
| collection | DOAJ |
| description | Deep learning methods have been successfully applied to image processing, mainly using 2D vision sensors. Recently, the rise of depth cameras and other similar 3D sensors has opened the field for new perception techniques. Nevertheless, 3D convolutional neural networks perform slightly worse than other 3D deep learning methods, and even worse than their 2D version. In this paper, we propose to improve 3D deep learning results by transferring the pretrained weights learned in 2D networks to their corresponding 3D version. Using an industrial object recognition context, we have analyzed different combinations of 3D convolutional networks (VGG16, ResNet, Inception ResNet, and EfficientNet), comparing the recognition accuracy. The highest accuracy is obtained with EfficientNetB0 using extrusion with an accuracy of 0.9217, which gives comparable results to state-of-the art methods. We also observed that the transfer approach enabled to improve the accuracy of the Inception ResNet 3D version up to 18% with respect to the score of the 3D approach alone. |
| first_indexed | 2024-03-09T05:38:45Z |
| format | Article |
| id | doaj.art-237c77384cc640479d350610fdfefcd4 |
| institution | Directory Open Access Journal |
| issn | 1424-8220 |
| language | English |
| last_indexed | 2024-03-09T05:38:45Z |
| publishDate | 2021-02-01 |
| publisher | MDPI AG |
| record_format | Article |
| series | Sensors |
| spelling | doaj.art-237c77384cc640479d350610fdfefcd42023-12-03T12:26:24ZengMDPI AGSensors1424-82202021-02-01214107810.3390/s210410783D Convolutional Neural Networks Initialized from Pretrained 2D Convolutional Neural Networks for Classification of Industrial PartsIbon Merino0Jon Azpiazu1Anthony Remazeilles2Basilio Sierra3TECNALIA, Basque Research and Technology Alliance (BRTA), Mikeletegi Pasealekua 7, 20009 Donostia-San Sebastián, SpainTECNALIA, Basque Research and Technology Alliance (BRTA), Mikeletegi Pasealekua 7, 20009 Donostia-San Sebastián, SpainTECNALIA, Basque Research and Technology Alliance (BRTA), Mikeletegi Pasealekua 7, 20009 Donostia-San Sebastián, SpainRobotics and Autonomous Systems Group, Universidad del País Vasco/Euskal Herriko Unibertsitatea, 48940 Basque, SpainDeep learning methods have been successfully applied to image processing, mainly using 2D vision sensors. Recently, the rise of depth cameras and other similar 3D sensors has opened the field for new perception techniques. Nevertheless, 3D convolutional neural networks perform slightly worse than other 3D deep learning methods, and even worse than their 2D version. In this paper, we propose to improve 3D deep learning results by transferring the pretrained weights learned in 2D networks to their corresponding 3D version. Using an industrial object recognition context, we have analyzed different combinations of 3D convolutional networks (VGG16, ResNet, Inception ResNet, and EfficientNet), comparing the recognition accuracy. The highest accuracy is obtained with EfficientNetB0 using extrusion with an accuracy of 0.9217, which gives comparable results to state-of-the art methods. We also observed that the transfer approach enabled to improve the accuracy of the Inception ResNet 3D version up to 18% with respect to the score of the 3D approach alone.https://www.mdpi.com/1424-8220/21/4/1078computer visiondeep learningtransfer learningobject recognition |
| spellingShingle | Ibon Merino Jon Azpiazu Anthony Remazeilles Basilio Sierra 3D Convolutional Neural Networks Initialized from Pretrained 2D Convolutional Neural Networks for Classification of Industrial Parts Sensors computer vision deep learning transfer learning object recognition |
| title | 3D Convolutional Neural Networks Initialized from Pretrained 2D Convolutional Neural Networks for Classification of Industrial Parts |
| title_full | 3D Convolutional Neural Networks Initialized from Pretrained 2D Convolutional Neural Networks for Classification of Industrial Parts |
| title_fullStr | 3D Convolutional Neural Networks Initialized from Pretrained 2D Convolutional Neural Networks for Classification of Industrial Parts |
| title_full_unstemmed | 3D Convolutional Neural Networks Initialized from Pretrained 2D Convolutional Neural Networks for Classification of Industrial Parts |
| title_short | 3D Convolutional Neural Networks Initialized from Pretrained 2D Convolutional Neural Networks for Classification of Industrial Parts |
| title_sort | 3d convolutional neural networks initialized from pretrained 2d convolutional neural networks for classification of industrial parts |
| topic | computer vision deep learning transfer learning object recognition |
| url | https://www.mdpi.com/1424-8220/21/4/1078 |
| work_keys_str_mv | AT ibonmerino 3dconvolutionalneuralnetworksinitializedfrompretrained2dconvolutionalneuralnetworksforclassificationofindustrialparts AT jonazpiazu 3dconvolutionalneuralnetworksinitializedfrompretrained2dconvolutionalneuralnetworksforclassificationofindustrialparts AT anthonyremazeilles 3dconvolutionalneuralnetworksinitializedfrompretrained2dconvolutionalneuralnetworksforclassificationofindustrialparts AT basiliosierra 3dconvolutionalneuralnetworksinitializedfrompretrained2dconvolutionalneuralnetworksforclassificationofindustrialparts |