TY - JOUR

T1 - Industrial Fisheye Image Segmentation Using Neural Networks

AU - Beloshapko, A.

AU - Korkhov, V.

AU - Knoll, C.

AU - Iben, U.

N1 - Conference code: 19

PY - 2019/7/1

Y1 - 2019/7/1

N2 - Fisheye cameras have recently became very popular in computer vision applications due to their wide field of view. In addition to a better overview of the surrounding area, they enable to capture objects at extremely close ranges. These advantages come at a cost of strong image distortion, which cannot be removed completely maintaining image continuity. This complicates the use of traditional computer vision algorithms, which expect a single image as an input. This paper presents a performance evaluation of neural network algorithms for object detection and segmentation on fisheye camera images. Three approaches are evaluated: semantic image segmentation with Fully Convolutional Network (FCN) [13], a fully convolutional approach to instance segmentation with U-Net [18] and a region-based approach to instance segmentation with Mask R-CNN [10]. All of these networks successfully solved the task. However, as they were designed to different purposes, each of them has its own strengths and shortcomings. These three approaches are used to perform euro container image segmentation task. An image dataset was created in order to train and evaluate these algorithms. Huge part of this dataset was generated artificially, which simplified the task of ground truth labeling. The power of neural networks enable for fast and reliable image segmentation. As to our knowledge, this is the first neural networks application for euro container fisheye image detection and segmentation.

AB - Fisheye cameras have recently became very popular in computer vision applications due to their wide field of view. In addition to a better overview of the surrounding area, they enable to capture objects at extremely close ranges. These advantages come at a cost of strong image distortion, which cannot be removed completely maintaining image continuity. This complicates the use of traditional computer vision algorithms, which expect a single image as an input. This paper presents a performance evaluation of neural network algorithms for object detection and segmentation on fisheye camera images. Three approaches are evaluated: semantic image segmentation with Fully Convolutional Network (FCN) [13], a fully convolutional approach to instance segmentation with U-Net [18] and a region-based approach to instance segmentation with Mask R-CNN [10]. All of these networks successfully solved the task. However, as they were designed to different purposes, each of them has its own strengths and shortcomings. These three approaches are used to perform euro container image segmentation task. An image dataset was created in order to train and evaluate these algorithms. Huge part of this dataset was generated artificially, which simplified the task of ground truth labeling. The power of neural networks enable for fast and reliable image segmentation. As to our knowledge, this is the first neural networks application for euro container fisheye image detection and segmentation.

KW - Euro container detection

KW - FCN

KW - Fisheye cameras

KW - Image instance segmentation

KW - Image semantic segmentation

KW - Mask R-CNN

KW - U-Net

UR - http://www.scopus.com/inward/record.url?scp=85068611843&partnerID=8YFLogxK

U2 - 10.1007/978-3-030-24305-0_50

DO - 10.1007/978-3-030-24305-0_50

M3 - Article

AN - SCOPUS:85068611843

VL - 11622

SP - 678

EP - 690

JO - Lecture Notes in Computer Science

JF - Lecture Notes in Computer Science

SN - 0302-9743

T2 - 19th International Conference on Computational Science and Its Applications, ICCSA 2019

Y2 - 1 July 2019 through 4 July 2019

ER -