De Mello, Alexandre R
(SENAI Innovation Institute for Embedded Systems, Florianó)
,
Barbosa, Flávio G O
(polis, Brazil)
,
Fonseca, Murilo L
(SENAI Innovation Institute for Embedded Systems, Florianó)
,
Smiderle, Camila D
(polis, Brazil)
This work aims at developing a solution to aid concrete dam inspections using unmanned aerial vehicle (UAV) imagery and DCNN-based (deep convolutional neural network) object detection. This paper introduces materials used to create the Sentinel dataset, which is a heterogeneous dataset that contains...
This work aims at developing a solution to aid concrete dam inspections using unmanned aerial vehicle (UAV) imagery and DCNN-based (deep convolutional neural network) object detection. This paper introduces materials used to create the Sentinel dataset, which is a heterogeneous dataset that contains 304 images on the training set and 316 images on the test set from a roller-compacted concrete water dam, with a Creager profile spillway, with annotation of three different types of objects: unwanted objects, harmless objects, and structural damage. This work proposes the use of the Faster R-RCNN and the Single Shot Multibox Detector, which are two deep convolutional neural networks for object detection with singular architectures, to identify nonconformity during an automatic dam inspection system, and analyzes the robustness of both networks considering the accuracy, true positive rate, true negative rate, precision, and F1 metrics. Using the Sentinel dataset, the best result yielded a Faster R-RCNN model with accuracy and an F1 score of 88.9%, which shows the viability of using the proposed solution in an operational environment.
This work aims at developing a solution to aid concrete dam inspections using unmanned aerial vehicle (UAV) imagery and DCNN-based (deep convolutional neural network) object detection. This paper introduces materials used to create the Sentinel dataset, which is a heterogeneous dataset that contains 304 images on the training set and 316 images on the test set from a roller-compacted concrete water dam, with a Creager profile spillway, with annotation of three different types of objects: unwanted objects, harmless objects, and structural damage. This work proposes the use of the Faster R-RCNN and the Single Shot Multibox Detector, which are two deep convolutional neural networks for object detection with singular architectures, to identify nonconformity during an automatic dam inspection system, and analyzes the robustness of both networks considering the accuracy, true positive rate, true negative rate, precision, and F1 metrics. Using the Sentinel dataset, the best result yielded a Faster R-RCNN model with accuracy and an F1 score of 88.9%, which shows the viability of using the proposed solution in an operational environment.
※ AI-Helper는 부적절한 답변을 할 수 있습니다.