Open AccessIMAGE LABELING FOR LIDAR INTENSITY IMAGE USING K-NN OF FEATURE OBTAINED BY CONVOLUTIONAL NEURAL NETWORK
Author(s) -
Masaki Umemura,
Kinichi Hotta,
Hidemi aka,
Kin-ya Oda
Publication year - 2016
Publication title -
the international archives of the photogrammetry, remote sensing and spatial information sciences/international archives of the photogrammetry, remote sensing and spatial information sciences
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.264
H-Index - 71
eISSN - 1682-1777
pISSN - 1682-1750
DOI - 10.5194/isprsarchives-xli-b3-931-2016
Subject(s) - artificial intelligence , pattern recognition (psychology) , feature (linguistics) , computer science , convolutional neural network , pixel , similarity (geometry) , ground truth , image (mathematics) , computer vision , k nearest neighbors algorithm , standard test image , feature extraction , image processing , philosophy , linguistics
We propose an image labeling method for LIDAR intensity image obtained by Mobile Mapping System (MMS) using K-Nearest Neighbor (KNN) of feature obtained by Convolutional Neural Network (CNN). Image labeling assigns labels (e.g., road, cross-walk and road shoulder) to semantic regions in an image. Since CNN is effective for various image recognition tasks, we try to use the feature of CNN (Caffenet) pre-trained by ImageNet. We use 4,096-dimensional feature at fc7 layer in the Caffenet as the descriptor of a region because the feature at fc7 layer has effective information for object classification. We extract the feature by the Caffenet from regions cropped from images. Since the similarity between features reflects the similarity of contents of regions, we can select top K similar regions cropped from training samples with a test region. Since regions in training images have manually-annotated ground truth labels, we vote the labels attached to top K similar regions to the test region. The class label with the maximum vote is assigned to each pixel in the test image. In experiments, we use 36 LIDAR intensity images with ground truth labels. We divide 36 images into training (28 images) and test sets (8 images). We use class average accuracy and pixel-wise accuracy as evaluation measures. Our method was able to assign the same label as human beings in 97.8% of the pixels in test LIDAR intensity images.