logo SBA

ETD

Archivio digitale delle tesi discusse presso l’Università di Pisa

Tesi etd-06212022-125535


Tipo di tesi
Tesi di laurea magistrale
Autore
GENTILI, ALESSANDRO
URN
etd-06212022-125535
Titolo
RAMI competition: detection and classification of man-made objects for the autonomy of underwater robots
Dipartimento
INGEGNERIA DELL'INFORMAZIONE
Corso di studi
INGEGNERIA ROBOTICA E DELL'AUTOMAZIONE
Relatori
relatore Prof. Costanzi, Riccardo
relatore Prof. Caiti, Andrea
relatore Dott. Bresciani, Matteo
Parole chiave
  • classification
  • cnn
  • competition
  • detection
  • digit
  • european
  • faster
  • image
  • man-made
  • network
  • neural
  • objects
  • preprocessing
  • RAMI
  • rcnn
  • recognition
  • underwater
Data inizio appello
07/07/2022
Consultabilità
Non consultabile
Data di rilascio
07/07/2092
Riassunto
The Metrological Evaluation and Testing of Robots in International CompetitionS (METRICS) EU project, led by the National Metrology and Testing Laboratory (LNE), organizes robotics competitions in four priority areas identified by the European Commission: health, agri-food, inspection and maintenance of infrastructure and agile production. In particular, the first edition of the Robotics for Asset Maintenance and Inspection (RAMI) competition will take place and will be led by NATO Science and Technology Organization - Centre for Maritime Research and Experimentation (STO CMRE), one of the reference research centres in Europe devoted to marine robotics technologies. RAMI Cascade competition will focus on the classification, identification and localization of Object of Potential Interest (OPI) in underwater images.

In this thesis work an object detection and classification algorithm able, to deal with underwater imaging and to detect and classify different OPIs, is developed. The training dataset containing OPI images employed during the preparation of the algorithm is the one proposed by NATO STO CMRE to the participating teams as support to the development of software. The provided images have been collected in underwater environment and divided into five classes: colored buoys, black numbers on yellow pipes, black numbers over red background, red markers on yellow pipes and images without OPIs.

The algorithm has been developed according to the following main steps. Firstly, an effective color enhancement and color restoration procedure has been defined and implemented on the basis of the Jaffe-McGlamery computer image degradation model and by employing pixel intensity redistribution techniques. Red and Blue channels have been pre-compensated to balance the high attenuation suffered in underwater images; then Gray World Assumption (GWA), Dark Channel Prior (DCP) and Contrast Limited Adaptive Histogram Equalization (CLAHE) have been implemented and applied to the RGB images provided in the dataset to compensate the color cast and to increase the overall image contrast. After an initial phase of this work in which a color and shape based detection and classification algorithm has been developed and tested, a novel approach based on deep learning is proposed. A state-of-the-art Faster R-CNN network has been fine-tuned and tested on detection and classification of three OPIs classes: buoys, black digits (1-6) and red markers. Here, data augmentation techniques and Early Stopping have been employed to increase the performances of the network and to prevent over-fitting. For the detection of the black digits (1-6), the patch provided as output by the Faster R-CNN network has been processed employing a binarization based on the Otsu's adaptive thresholding method. Then, a Convolutional Neural Network (CNN), trained on the modified National Institute of Standards and Technology (MNIST) dataset and fine-tuned on a dataset created exploiting all the ground-truth boxes containing black digits, has been used to recognize and classify each digit. At last, the performance of the deep learning based novel approach has been evaluated through mean Average Precision (mAP), class Average Precision (AP) and accuracy metrics. The results reported in the last part of this work, show that the first stage of the algorithm, based on the Faster R-CNN network, achieved a mean Average Precision (mAP) of about 0.94, while the Digit Recognition CNN reached an accuracy of about 77%.
File