• Journal of the Korea Institute of Military Science and Technology
• /
• v.24 no.4
• /
• pp.382-391
• /
• 2021

Unsupervised domain adaptation often gives impressive solutions to handle domain shift of data. Most of current approaches assume that unlabeled target data to train is abundant. This assumption is not always true in practices. To tackle this issue, we propose a general solution to solve the domain gap minimization problem without any target data. Our method consists of two regularization steps. The first step is a pixel regularization by arbitrary style transfer. Recently, some methods bring style transfer algorithms to domain adaptation and domain generalization process. They use style transfer algorithms to remove texture bias in source domain data. We also use style transfer algorithms for removing texture bias, but our method depends on neither domain adaptation nor domain generalization paradigm. The second regularization step is a feature regularization by feature alignment. Adding a feature alignment loss term to the model loss, the model learns domain invariant representation more efficiently. We evaluate our regularization methods from several experiments both on small dataset and large dataset. From the experiments, we show that our model can learn domain invariant representation as much as unsupervised domain adaptation methods.

• Proceedings of the IEEK Conference
• /
• 2000.06c
• /
• pp.157-160
• /
• 2000

This paper proposes a color-based image retrieval method using color adjacency and gradient. In proposed method, both the adjacency of different colors and gradient of a color in homogeneous region are considered as features of an image. The gradient, defined as the maximum distance along the direction with largest change of color, is computed for each pixel to determine whether the center color is similar or different to the neighboring colors. Therefore the problems caused by uniform quantization, which is popularly used at most existing retrieval, can be avoided. And furthermore, the storage of the feature is reduced by the proposed binary representation.

• Journal of the Korean Institute of Telematics and Electronics S
• /
• v.34S no.5
• /
• pp.35-44
• /
• 1997

In this paper, a an algorithm that detects the endocardial boundary, expanding the region from endocardial cavity using fuzzy inference, is proposed. This algorithm decides the ventricular cavity by fuzzy inference in process of searching each pixel from the inside of left ventricle in echocardial image and expands it. Uncertainty and fuzziness exists in decision of endocardial boundary. Therefore, we convert the lingustic representation of mean, standard deviation, and threshold value that are characteristic variables of endocardial boundary to fuzzy input and output variables. And, we extract proposed method is robuster to noise than radial searching method that is highly dependent on center position. To prove the similarity of detected boundary by fuzzy nference, we used the measures of SIZE, correlation coefficient, MSD, and RMSE and had acquired reasonable results.

• The Journal of the Acoustical Society of Korea
• /
• v.11 no.1
• /
• pp.42-51
• /
• 1992

This paper proposes an acquisition and the representation method of the 3-dimensional information. The proposed range finder system can reduce the computation time by only calculating the ${\triangle}R$ of each pixel compared to the existing methods. We also propose a shape feature extraction method by considering the sign change of the acquired range data. Finally, the effectiveness of this system is demonstrated by several experiments.

• Journal of Institute of Control, Robotics and Systems
• /
• v.8 no.6
• /
• pp.455-461
• /
• 2002

Object-based image retrieval method is addressed. A new image segmentation algorithm and image comparing method between segmented objects are proposed. For image segmentation, color and texture features are extracted from each pixel in the image. These features we used as inputs into VQ (Vector Quantization) clustering method, which yields homogeneous objects in terns of color and texture. In this procedure, colors are quantized into a few dominant colors for simple representation and efficient retrieval. In retrieval case, two comparing schemes are proposed. Comparing between one query object and multi objects of a database image and comparing between multi query objects and multi objects of a database image are proposed. For fast retrieval, dominant object colors are key-indexed into database.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.9 no.11
• /
• pp.4556-4572
• /
• 2015

Dynamic texture (DT) recognition is a challenging problem in numerous applications. In this study, we propose a new algorithm for DT recognition based on group sparsity structure in conjunction with chaotic feature vector. Bag-of-words model is used to represent each video as a histogram of the chaotic feature vector, which is proposed to capture self-similarity property of the pixel intensity series. The recognition problem is then cast to a group sparsity model, which can be efficiently optimized through alternating direction method of multiplier algorithm. Experimental results show that the proposed method exhibited the best performance among several well-known DT modeling techniques.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 1999.06a
• /
• pp.77-82
• /
• 1999

Information about camera operating such as zoom, focus, pan, tilt and tracking is useful not only for efficient video coding, but also for content-based video representation. A camera operating parameter measurement system designed specifically for these applications is therefore developed. This system, implemented in real time and synchronized with the video signal, measures the precise camera operating parameters. We calibrated the camera lens using a camera model that accounts for redial lens distortion. The system is then applied to infer image motion from pan and tilt operating parameters. The experimental results show that the inferred motion coincides with the actual motion very well, with an error of less than 0.5 pixel even for large motion up to 80 pixels.

• Proceedings of the KSRS Conference
• /
• 2003.11a
• /
• pp.645-647
• /
• 2003

In this work a new method for on-line scene segmentation is developed. In remote sensing a scene is represented by the pixel-oriented features. It is possible to reduce data redundancy by an unsupervised segment-feature extraction process, where the segment-features, rather than the pixelfeatures, are used for multispectral scene representation. The algorithm partitions the observation space into exhaustive set of disjoint segments. Then, pixels belonging to each segment are characterized by segment features. Illustrative examples are presented, and the performance of features is investigated. Results show an average compression more than 25, the classification performance is improved for all classes, and the CPU time required for classification is reduced by the same factor.

• Transactions of the Korean Society of Mechanical Engineers A
• /
• v.40 no.4
• /
• pp.373-379
• /
• 2016

It is known that it is expedient to represent the distribution of the properties of a bone with complex heterogeneity as B-spline volume functions. For B-spline-based representation, the pixel values of CT images are interpolated by B-spline volume functions. However, the CT images of a bone are three-dimensional and very large, and hence a large amount of memory and long computation time for the interpolation are required. In this study, a method for resolving these problems is proposed. In the proposed method, the B-spline volume interpolation problem is simplified by using the uniformity of pixel spacing of the image and the properties of B-spline basis functions. This results in a reduction in computation time and the amount of memory used. The proposed method was implemented and it was verified that the computation time and the amount of memory used were reduced.

• The Journal of Korea Institute of Information, Electronics, and Communication Technology
• /
• v.11 no.3
• /
• pp.229-234
• /
• 2018

Texture is one of the most useful features in classifying and segmenting images. The LBP-based approach previously presented in the literature has been successful in many applications. However, it's theoretical foundation is based only on the difference of pixel values, and consequently it has a number of drawbacks like it performs poorly for the images corrupted with noise, and especially it cannot be used as a multiscale texture descriptor due to the exploding increase of feature vector dimension with increase of the number of neighbor pixels. In this paper, we present a method to address these drawbacks of LBP-based approach. More specifically, our approach quantizes the range of pixels values and construct a 3D histogram which captures the correlative information of pixels. This histogram is used as a texture feature. Several tests with texture images show that the proposed method outperforms the LBP-based approach in the problem of texture classification.