Convolutional neural networks for recognition and segmentation of aluminium profiles

The performance of the proposed system strongly depends on the dataset used for training and on the backbone architecture used. This why, adopting a strategy that generates a priori a very large dataset the time required to create the annotation file grows almost exponentially. The validation and test dataset, on the other hand, consists of real images captured by cameras. The distinctiveness of this work consists to train a deep neural network with synthetic data (aluminium CAD files) and verify if it on real data. Experiments show that the implementation of architecture, as described above, leads to good performance in automatic detection and classification. Future work will be addressed to improve the network training process together with the architecture, the algorithms and the dataset creation process. The latter is proved to be fundamental for the balance and optimization of the whole process. The way is to develop not much augmented datasets focusing on online data augmentation during network training.