• Journal of IKEEE
• /
• v.28 no.1
• /
• pp.65-71
• /
• 2024

This paper proposes a neural network that can reduce the number of parameters while reducing matching errors in occluded regions to increase the accuracy of depth maps in stereo matching. Stereo matching-based object recognition is utilized in many fields to more accurately recognize situations using images. When there are many objects in a complex image, an occluded area is generated due to overlap between objects and occlusion by background, thereby lowering the accuracy of the depth map. To solve this problem, existing research methods that create context information and combine it with the cost volume or RoIselect in the occluded area increase the complexity of neural networks, making it difficult to learn and expensive to implement. In this paper, we create a depthwise seperable neural network that enhances regional feature extraction before cost volume generation, reducing the number of parameters and proposing a neural network that is robust to occlusion errors. Compared to PSMNet, the proposed neural network reduced the number of parameters by 30%, improving 5.3% in color error and 3.6% in test loss.

• Proceedings of the KSRS Conference
• /
• 2002.10a
• /
• pp.74-74
• /
• 2002

This paper describes a corner matching technique for automatic relative orientation. Automatically matched corner points from stereo aerial images are used to a data set and help to improve automation of relative orientation process. A general corner matching process of overall image to image has very heavy operation and repetitive computation, so very time-consuming. But aerial stereo images are approximately seventy percent overlapped and little rotated. Based this hypothesis, we designed a sectional corner matching technique calculating correlation section by section between stereo images. Although the overlap information is not accurate, if we know it approximately, the matching process can be lighter. Since the size of aerial image is very large, corner extraction process is performed hierarchically by creating image pyramid, and corners extracted are refined at the higher level image. Extracted corners at the final step are matched section by section. Matched corners are filtered using positional information and their relation and distribution. Filtering process is applied over several steps because the thing affecting to get good result-good relative orientation parameter- is not the number of matched corner points but the accuracy of them. Filtered data is filtered one more during the process calculating relative orientation parameters. When the process is finished, we can get the well matched corner points and the refined Von-Gruber areas besides relative orientation parameters. This sectional corner matching technique is efficient by decreasing unnecessarily repetitive operations and contributes to improve automation for relative orientation.

• The KIPS Transactions:PartB
• /
• v.9B no.2
• /
• pp.237-244
• /
• 2002

In this paper, we propose a stereo image recognition-based method for measuring the volume of the rectangular parallelepiped. The method measures the volume from two images captured with two CCD (charge coupled device) cameras by sequential processes such as ROI (region of interest) extraction, feature extraction, and stereo matching-based vortex recognition. The proposed method makes it possible to measure the volume of the 3D object at high speed because only a few features are used in the process of stereo matching. From experimental results, it is demonstrated that this method is very effective for measuring the volume of the rectangular parallelepiped at high speed.

• Journal of Internet Computing and Services
• /
• v.10 no.4
• /
• pp.25-33
• /
• 2009

In the paper, we proposed a new system to retrieve 3D wrinkle data based on stereo images. Usually, 3D reconstruction based on stereo images or video is very popular and it is the research focus, which has been applied for culture heritage, building and other scene. The target is object measurement, the scene depth calculation and 3D data obtained. There are several challenges in our research. First, it is hard to take the full information wrinkle images by cameras because of light influence, skin with non-rigid object and camera performance. We design a particular computer vision system to take winkle images with a long length camera lens. Second, it is difficult to get the dense stereo data because of the hard skin texture image segmentation and corner detection. We focus on semi-dense stereo matching algorithm for the wrinkle depth. Compared with the 3D scanner, our system is much cheaper and compared with the physical modeling based method, our system is more flexible with high performance.

• Journal of the Korean Society of Surveying, Geodesy, Photogrammetry and Cartography
• /
• v.16 no.2
• /
• pp.177-185
• /
• 1998

There is an increasing request for 3D data and outlines on building for urban planning and design. This paper describes an approach to extract building using Digital Surface Models(DSM) and stereo pairs of aerial images. DSM contain informations not only about the topographic surface like Digital Elevation Models(DEM), but also about buildings and other objects higher than the surrounding topographic surface, e.g. tees. We therefore describe our approach consisting of two step procedures. The first step of the approach is to generate DSM by stereo matching using Maximum Likelihood Estimation and Dynamic Programming. The proposed stereo matching is using the cost function for finding the disparity between the left and right image, and the Dynamic Programming for solving the stereo matching problem. The second step is to detect building outlines using the DSM and the edge informations extracted from a digital aerial image by Sobel Operator. The overlay analysis of the DSM and the edge information by Sobel Operator was efficient to detect building outlines.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.28 no.4C
• /
• pp.424-432
• /
• 2003

In this paper, an effective method of stereo image reconstruction through the regularized adaptive disparity estimation is proposed. Althougth the conventional adaptive disparity estimation method can sharply improve the PSNR of a reconstructed stereo image, but some problems of overlapping between the matching windows and disallocation of the matching windows can be occurred, because the matching window size changes adaptively in accordance with the magnitude of feature values. Accordingly, in thia paper, a new regularized adaptive disparity estimation technique is proposed. That is, by regularizing the estimated disparity vector with the neughboring disparity vectors, problems of the conventional adaptive disparity estimated scheme might be solved, and also the predicted stereo image can be more effectively reconstructed. From some experiments using the CCETT'S stereo image pairs of 'Man' and 'Claude', it is analyzed that the proposed disparity estimation scheme can improve PSNRs of the reconstructed images to 10.89dB, 6.13dB for 'Man' and 1.41dB, 0.81dB for 'Claude' by comparing with those of the conventional pixel-based and adaptive estimation method, respectively.

• Korean Journal of Remote Sensing
• /
• v.11 no.3
• /
• pp.61-70
• /
• 1995

In this paper, the algorithm to extract the height of flat-rooftop structures in stereo aerial image is suggested with an assumption that location, orientation, focal length, and field of view of a camera are known. It can be adapted to stereo aerial or satellite images. For performing feature-based stereo matching, the line segments suitable to describe the shape of general buildings are chosen as the feature. This paper is composed of three categories;the first step is to extract edges of structures with the polygon extraction algorithm which utilizes the edge following method, the second step is to perform the line segment matching with the camera information, and the last step is to calculate the location of each matched line and to estimate heights. The stereo images used in experiments are not real but synthetic ones. The experiment shows good results.

• Convergence Security Journal
• /
• v.13 no.3
• /
• pp.9-15
• /
• 2013

Calculating of matching cost is an important for efficient stereo matching. To investigate the performance of matching process, the concepts of the existing methods are introduced. Also we analyze the performance and merits of them. The simplest matching costs assume constant intensities at matching image locations. We consider matching cost functions which can be distinguished between pixel-based and window-based approaches. The Pixel-based approach includes absolute differences (AD) and sampling-intensitive absolute differences (BT). The window-based approach includes the sum of the absolute differences, the sum of squared differences, the normalized cross-correlation, zero-mean normalized cross-correlation, census transform, and the absolute differences census transform (AD-Census). We evaluate matching cost functions in terms of accuracy and time complexity. In terms of the accuracy, AD-Census method shows the lowest matching error ratio (the best solution). The ZNCC method shows the lowest matching error ratio in non-occlusion and all evaluation part. But it performs high matching error ratio at the discontinuities evaluation part due to blurring effect in the boundary. The pixel-based AD method shows a low complexity in terms of time complexity.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2003.05a
• /
• pp.171-174
• /
• 2003

This paper is a study on the 3D infrared image enhancement with Stereoscopic algorithm on still infrared image. The adapted stereo method is that the depth is extracted by comparison with right-left image, and the enhanced 3D infrared image by matching based on feature is realized. As the result of experiment this method forced the more smooth edge lines of 3D infrared images.

• Proceedings of the KIEE Conference
• /
• 2005.07d
• /
• pp.2531-2533
• /
• 2005

The proposed neural network technique is the real time computation method based theory of inter-node diffusion for searching the safety distances from the sudden appearance-objects during the work driving. The main steps of the distance computation using the theory of stereo vision like the eyes of man is following steps. One is the processing for finding the corresponding points of stereo images and the other is the interpolation processing of full image data from nonlinear image data of objects. All of therm request much memory space and time. Therefore the most reliable neural-network algorithm is drived for real-time matching of obejects, which is composed of a dynamic programming algorithm based on sequence matching techniques in moving objects.