• Journal of Multimedia Information System
• /
• v.9 no.3
• /
• pp.177-182
• /
• 2022

A deep neural network (DNN) includes variables whose values keep on changing with the training process until it reaches the final point of convergence. These variables are the co-efficient of a polynomial expression to relate to the feature extraction process. In general, DNNs work in multiple 'dimensions' depending upon the number of channels and batches accounted for training. However, after the execution of feature extraction and before entering the SoftMax or other classifier, there is a conversion of features from multiple N-dimensions to a single vector form, where 'N' represents the number of activation channels. This usually happens in a Fully connected layer (FCL) or a dense layer. This reduced 2D feature is the subject of study for our analysis. For this, we have used the FCL, so the trained weights of this FCL will be used for the weight-class correlation analysis. The popular DNN models selected for our study are ResNet-101, VGG-19, and GoogleNet. These models' weights are directly used for fine-tuning (with all trained weights initially transferred) and scratch trained (with no weights transferred). Then the comparison is done by plotting the graph of feature distribution and the final FCL weights.

• Journal of IKEEE
• /
• v.13 no.2
• /
• pp.181-185
• /
• 2009

In this paper, we are concerned with a 3D line segment extraction method by area-based stereo matching technique. The main idea is based on line fitting of elevation data on 2D line coordinates of ortho-image. Elevation data and ortho-image can be obtained by well-known area-based stereo matching technique. In order to use elevation in line fitting, the elevation itself should be reliable. To measure the reliability of elevation, in this paper, we employ the concept of self-consistency. We test the effectiveness of the proposed method with a quantitative accuracy analysis using synthetic images generated from Avenches data set of Ascona aerial images. Experimental results indicate that our method generates 3D line segments almost 7.5 times more accurate than raw elevations obtained by area-based method.

• ETRI Journal
• /
• v.33 no.4
• /
• pp.589-599
• /
• 2011

This paper presents an image description and matching scheme using synthetic features for a recommendation service. The recommendation service is an example of smart search because it offers something before a user's request. In the proposed extraction scheme, an image is described by synthesized spatial and statistical features. The spatial feature is designed to increase the discriminability by reflecting delicate variations. The statistical feature is designed to increase the robustness by absorbing small variations. For extracting spatial features, we partition the image into concentric circles and extract four characteristics using a spatial relation. To extract statistical features, we adapt three transforms into the image and compose a 3D histogram as the final statistical feature. The matching schemes are designed hierarchically using the proposed spatial and statistical features. The result shows that each feature is better than the compared algorithms that use spatial or statistical features. Additionally, if we adapt the proposed whole extraction and matching scheme, the overall performance will become 98.44% in terms of the correct search ratio.

• International Journal of CAD/CAM
• /
• v.6 no.1
• /
• pp.117-123
• /
• 2006

Recently with the development of 3D modeling and digitizing tools, more and more models have been created, which leads to the necessity of the technique of 3D mode retrieval system. In this paper we investigate a new method for 3D model retrieval based on orthogonal projections. We assume that 3D models are composed of trigonal meshes. Algorithms process first by a normalization step in which the 3D models are transformed into the canonical coordinates. Then each model is orthogonally projected onto six surfaces of the projected cube which contains it. A following step is feature extraction of the projected images which is done by Moment Invariants and Polar Radius Fourier Transform. The feature vector of each 3D model is composed of the features extracted from projected images with different weights. Our System validates that this means can distinguish 3D models effectively. Experiments show that our method performs quit well.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.11 no.3
• /
• pp.1595-1613
• /
• 2017

The IR camera and laser-based IR projector provide an effective solution for real-time collection of moving targets in RGB-D videos. Different from the traditional RGB videos, the captured depth videos are not affected by the illumination variation. In this paper, we propose a novel feature extraction framework to describe human activities based on the above optical video capturing method, namely spatial-temporal texture features for 3D human activity recognition. Spatial-temporal texture feature with depth information is insensitive to illumination and occlusions, and efficient for fine-motion description. The framework of our proposed algorithm begins with video acquisition based on laser projection, video preprocessing with visual background extraction and obtains spatial-temporal key images. Then, the texture features encoded from key images are used to generate discriminative features for human activity information. The experimental results based on the different databases and practical scenarios demonstrate the effectiveness of our proposed algorithm for the large-scale data sets.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2022.10a
• /
• pp.256-257
• /
• 2022

In this paper, we introduce a method of extracting three-dimensional feature points through the movement of a single mobile device. Using a monocular camera, a 2D image is acquired according to the camera movement and a baseline is estimated. Perform stereo matching based on feature points. A feature point and a descriptor are acquired, and the feature point is matched. Using the matched feature points, the disparity is calculated and a depth value is generated. The 3D feature point is updated according to the camera movement. Finally, the feature point is reset at the time of scene change by using scene change detection. Through the above process, an average of 73.5% of additional storage space can be secured in the key point database. By applying the algorithm proposed to the depth ground truth value of the TUM Dataset and the RGB image, it was confirmed that the\re was an average distance difference of 26.88mm compared with the 3D feature point result.

• The Transactions of the Korea Information Processing Society
• /
• v.7 no.5
• /
• pp.1361-1369
• /
• 2000

For the implementation of 3D Sound Localization system, the binaural filtering by HRTFs is generally employed. But the HRTF filter is of high order and its coefficients for all directions have to be stored, which imposes a rather large memory requirement. To cope with this, research works have centered on obtaining low dimensional HRTF representations without significant loss of information and synthesizing the original HRTF efficiently, by means of feature extraction methods for multivariate dat including PCA. In these researches, conventional linear PCA was applied to the frequency domain HRTF data and using relatively small number of principal components the original HRTFs could be synthesized in approximation. In this paper we applied neural network based nonlinear PCA model (NLPCA) and the nonlinear PLS repression model (NLPLS) for this low dimensional HRTF modeling and analyze the results in comparison with the PCA. The NLPCA that performs projection of data onto the nonlinear surfaces showed the capability of more efficient HRTF feature extraction than linear PCA and the NLPLS regression model that incorporates the direction information in feature extraction yielded more stable results in synthesizing general HRTFs not included in the model training.

• 제어로봇시스템학회:학술대회논문집
• /
• 1997.10a
• /
• pp.1848-1851
• /
• 1997

This paper presents the development of depth detection algorithm for the 3D Endoscopic Data using a stereo matchod and depth calcuation.

• Proceedings of the Korean Society of Precision Engineering Conference
• /
• 2013.05a
• /
• pp.919-920
• /
• 2013

• Journal of the Korean Society of Surveying, Geodesy, Photogrammetry and Cartography
• /
• v.21 no.2
• /
• pp.165-172
• /
• 2003

IKONOS satellite images are particularly well suited for stereo feature extraction. But, because IKONOS doesn't offer information about the satellite ephemeris and attitude, we have to use IKONOS RPC(Rational Polynomial Coefficients) data for 3-D feature extraction. In this study, it was intended to increase the accuracy and the efficiency in application of high resolution satellite images. Therefore, this study develop the program to extract 3-D feature information and have analyzed the geometric accuracy of the IKONOS satellite images by means of the change with the number, distribution and height of GCPs. This study will provide basic information for luther studies of the accuracy correction in IKONOS and high resolution satellite images.