Object Detection Model Training Framework for Very Small Datasets Applied to Outdoor Industrial Structures

No Thumbnail Available
Baharuddin M.Z.
How D.N.T.
Sahari K.S.M.
Abas A.Z.
Ramlee M.K.
Journal Title
Journal ISSN
Volume Title
Springer Science and Business Media Deutschland GmbH
Research Projects
Organizational Units
Journal Issue
Visual inspection of electrical utility assets is crucial in ensuring the continuous operation of a system or plant. With the advent of digital imagery using mobile devices, it has become easy to collect a vast amount of asset pictures from sites. To further enhance inspection efficiency, we propose RetinaNet, a deep learning-based object detection model that can be trained to automatically detect specific objects and features from images of outdoor industrial structures. The model is capable of detecting features such as intrusions, tree or bushes in the vicinity of the lattice towers. We also introduce a model training framework for use with very small datasets which consists of rigorous data augmentation, image pre-sizing, focal loss function, progressive resizing, learning rate finder, and the Ranger optimizer. Experiment results show that the proposed model used in conjunction with the aforementioned training framework results in the lowest validation loss and highest mean average precision of 31.36 � 2021, Springer Nature Switzerland AG.
Computer vision; Convolutional neural networks; Deep neural networks; Digital devices; Feature extraction; Image enhancement; Inspection; Object recognition; Convolutional neural network; Deep learning; Detection models; Industrial inspections; Industrial structures; Model training; Objects detection; Retinanet; Small data set; Training framework; Object detection