|
|
Fast and Effective Motion Model for Moving Object Detection Using Aerial Images
Zainal Rasyid Mahayuddin, University Kebangsaan Malaysia, Malaysia
A.F.M. Saifuddin Saif, American International University-Bangladesh
Abstract:
Motion detection remains an unsolved issue for moving object detection
using aerial images from moving vehicle like Unmanned Aerial Vehicle
due to lack of motion model. Existing moving object detection methods
do not provide motion model to detect motion pixels. In addition, previous research for moving object detection depends on either frame
difference or segmentation methods. Frame difference based approaches can
differentiate pixel motion but cannot extract the overall object whereas
segmentation approaches can extract the overall object but cannot differentiate object motion. However, moving object performance depends
on the feature type(s) employed, due to limited feature availability from
aerial images. The purpose of the current research is first to select a new
feature for overall detection procedures, second to propose a model for
motion detection, and third to apply frame difference and segmentation
methods together to achieve optimum detection performance. A new
motion model, Advanced Moment based Motion Unification (AMMU) is
proposed, where the moment feature is used for motion detection. Experimental results verify that the proposed AMMU model is successful
at detection of moving objects.
Full Paper (in PDF)
|
Ovarian Follicle Classification using Convolutional Neural Networks from Ultrasound Scanning Images
Manabu Nii, Yusuke Kato, Masakazu Morimoto, Syoji Kobashi, Naotake Kamiura,
Graduate School of Engineering, University of Hyogo, Himeji, Hyogo, Japan
Yutaka Hata, Graduate School of Simulation Studies, University of Hyogo, Kobe, Hyogo, Japan
Seturo Imawaki, Ishikawa Hospital, Hyogo, Japan
Tomomoto Ishikawa, Hidehiko Matsubayashi, Reproduction Clinic Osaka, Osaka, Japan
Abstract:
In this paper, a new approach to classify ovarian follicles into two classes is proposed.
A smoothing filter is applied for filtering ovarian follicle images.
The smoothing filter is designed to consider speckle patterns under the resolution of the ultrasound devices.
Then, for extracting features from the filtered ovarian follicle images, two types of convolutional neural networks are utilized.
One is the convolutional autoencoder, and the other is the layered convolutional neural network.
Finally, both features extracted by the CNN-AE or the CNN from the filtered ovarian follicle images and numerical features defined by our previous works are used for classification.
Several types of classifiers are examined in our experiments.
From experimental results, we show the effectiveness of our proposed method.
Especially, when image features extracted by the CNN and numerical features are both used, we have better classification performance than the other cases.
Full Paper (in PDF)
|
|
