Earticle

The purpose of this study was to develop and evaluate Point Cloud Data (PCD) deep learning models and a rule-based system for segmenting tree structures (stems and crowns) using fixed terrestrial LiDAR data. The dataset comprised 48 Larix Kaemferi trees, which were collected and preprocessed. For the PCD deep learning models, three downsampled datasets consisting of 1024, 4096, and 16384 points were constructed from the original data. The data was divided into training (70%) and validation (30%) sets. Models were built using PointNet and PointNet++ architectures, resulting in a total of 12 tree structure segmentation models for accuracy comparison. The rule-based system was developed using the original data, applying techniques such as verticality checks, cylindrical structure detection, and slice-based circular fitting to detect the stem. It then segmented the stem through repetitive circle fitting and validation processes based on height. The average accuracy of the PCD deep learning tree structure segmentation models was approximately 95%, with the PointNet++ model using 16384 points achieving the highest classification accuracy of about 98%. The rule-based system achieved high classification accuracy of over 99% for both tree species. This study is expected to contribute to precise measurement and efficient management of forest resources by presenting automated methods for tree structure segmentation using AI technology and rule-based approaches. It is anticipated that this research will serve as a foundation for the advancement of forest digitalization, precision forest management technologies, forest structure analysis, and timber production estimation in various fields.

This research aimed to assess the possibility of detecting forest degradation using time-series satellite imagery and three different deep learning-based change detection techniques. The dataset used for the deep learning models was composed of two sets, one based on surface reflectance (SR) spectral information from satellite imagery, combined with Texture Information (GLCM; Gray-Level Co-occurrence Matrix) and terrain information. The deep learning models employed for land cover change detection included image differencing using the Unet semantic segmentation model, multi-encoder Unet model, and multi-encoder Unet++ model. The study found that there was no significant difference in accuracy between the deep learning models for forest degradation detection. Both training and validation accuracies were approximately 89% and 92%, respectively. Among the three deep learning models, the multi-encoder Unet model showed the most efficient analysis time and comparable accuracy. Moreover, models that incorporated both texture and gradient information in addition to spectral information were found to have a higher classification accuracy compared to models that used only spectral information. Overall, the accuracy of forest degradation extraction was outstanding, achieving 98%.

컴퓨터 비전 분야의 의미론적 영상 분할(Semantic Image Segmentation) 기술은 이미지를 픽셀 단위로 분할 하여 클래스를 나누는 기술이다. 이 기술도 기계 학습을 이용한 방법으로 성능이 빠르게 향상되는 중이며, 픽셀 단위의 정보를 활용할 수 있는 높은 활용성이 주목받는 기술이다. 그러나 이 기술은 초기부터 최근까지도 계속 '세밀하지 못한 분할'에 대한 문제가 제기되어 왔다. 이 문제는 레이블 맵의 크기를 계속 늘리면서 발생한 문제이기 때문에, 자세한 에지 정보가 있는 원본 영상의 에지 맵을 이용해 레이블 맵을 수정하여 개선할 수 있을 것으로 예상할 수 있었다. 따라서 본 논문은 기존 방법대로 학습 기반의 의미론적 영상 분할을 유지하되, 그 결과인 레이블 맵을 원본 영상의 에지 맵 기반으로 수정하는 후처리 알고리즘을 제안한다. 기존의 방법에 알고리즘의 적용 한 뒤 전후의 정확도를 비교했을 때 평균적으로 약 1.74% 픽셀 정확도와 1.35%의 IoU(Intersection of Union) 정확도가 향상되었으며, 결과를 분석했을 때 성공적으로 본래 목표한 세밀한 분할 기능을 개선했음을 보였다.

Semantic image segmentation technology in the field of computer vision is a technology that classifies an image by dividing it into pixels. This technique is also rapidly improving performance using a machine learning method, and a high possibility of utilizing information in units of pixels is drawing attention. However, this technology has been raised from the early days until recently for 'lack of detailed segmentation' problem. Since this problem was caused by increasing the size of the label map, it was expected that the label map could be improved by using the edge map of the original image with detailed edge information. Therefore, in this paper, we propose a post-processing algorithm that maintains semantic image segmentation based on learning, but modifies the resulting label map based on the edge map of the original image. After applying the algorithm to the existing method, when comparing similar applications before and after, approximately 1.74% pixels and 1.35% IoU (Intersection of Union) were applied, and when analyzing the results, the precise targeting fine segmentation function was improved.

This paper proposes Multidirectional Block Ranking based segmentation and removal of Interphase cells from chromosome images. The efficiency of automatic karyotyping decreases with the presence of undivided, condensed mass of chromosomes called Interphase cells, stain, debris and other unwanted interferences in the chromosome image. The algorithm segments and removes these interferences and enhances the accuracy of automatic karyotyping. The method is tested and excellent segmentation accuracy is accomplished. The chromosome image is preprocessed and a boundary-mapping algorithm is applied to identify the Region of Interest (RoI). The image is divided into blocks and ranks are assigned to all the blocks using Gaussian Ranking Functions (GRF) based on their contribution to the RoI. A higher rank is assigned to the block contributing more while a lesser rank is assigned to other blocks that contribute less to the RoI. The Interphase cells that constitute the RoI are removed from the chromosome images based on the cumulative rank obtained by the blocks in the chromosome image. The proposed algorithm is applied to segment and remove Interphase cells, stains, dirt and other particles that exhibit structural homogeneity. The algorithm gives accurate results in applications where the RoI to be segmented share the grey level with the background wherein the traditional image segmentation methods fall short of accomplishing precise segmentation results.

In this paper, layered UNet with warmup and dropout tricks was used to segment teeth instantly by using data labeled for each individual tooth and increase performance of the result. The layered UNet proposed before showed very good performance in tooth segmentation without distinguishing tooth number. To do instance segmentation of teeth, we labeled teeth CBCT data according to tooth numbering system which is devised by FDI World Dental Federation notation. Colors for labeled teeth are like AI-Hub teeth dataset. Simulation results show that layered UNet does also segment very well for each tooth distinguishing tooth number by color. Layered UNet model using warmup trick was the best with IoU values of 0.80 and 0.77 for training, validation data. To increase the performance of instance segmentation of teeth, we need more labeled data later. The results of this paper can be used to develop medical software that requires tooth recognition, such as orthodontic treatment, wisdom tooth extraction, and implant surgery.

Medical image segmentation is the most important task in radiation therapy. Especially, when segmenting medical images, the liver is one of the most difficult organs to segment because it has various shapes and is close to other organs. Therefore, automatic segmentation of the liver in computed tomography (CT) images is a difficult task. Since tumors also have low contrast in surrounding tissues, and the shape, location, size, and number of tumors vary from patient to patient, accurate tumor segmentation takes a long time. In this study, we propose a method algorithm for automatically segmenting the liver and tumor for this purpose. As an advantage of setting the boundaries of the tumor, the liver and tumor were automatically segmented from the CT image using the 2D CoordConv DeepLab V3+ model using the CoordConv layer. For tumors, only cropped liver images were used to improve accuracy. Additionally, to increase the segmentation accuracy, augmentation, preprocess, loss function, and hyperparameter were used to find optimal values. We compared the CoordConv DeepLab v3+ model using the CoordConv layer and the DeepLab V3+ model without the CoordConv layer to determine whether they affected the segmentation accuracy. The data sets used included 131 hepatic tumor segmentation (LiTS) challenge data sets (100 train sets, 16 validation sets, and 15 test sets). Additional learned data were tested using 15 clinical data from Seoul St. Mary's Hospital. The evaluation was compared with the study results learned with a two-dimensional deep learning-based model. Dice values without the CoordConv layer achieved 0.965 ± 0.01 for liver segmentation and 0.925 ± 0.04 for tumor segmentation using the LiTS data set. Results from the clinical data set achieved 0.927 ± 0.02 for liver division and 0.903 ± 0.05 for tumor division. The dice values using the CoordConv layer achieved 0.989 ± 0.02 for liver segmentation and 0.937 ± 0.07 for tumor segmentation using the LiTS data set. Results from the clinical data set achieved 0.944 ± 0.02 for liver division and 0.916 ± 0.18 for tumor division. The use of CoordConv layers improves the segmentation accuracy. The highest of the most recently published values were 0.960 and 0.749 for liver and tumor division, respectively. However, better performance was achieved with 0.989 and 0.937 results for liver and tumor, which would have been used with the algorithm proposed in this study. The algorithm proposed in this study can play a useful role in treatment planning by improving contouring accuracy and reducing time when segmentation evaluation of liver and tumor is performed. And accurate identification of liver anatomy in medical imaging applications, such as surgical planning, as well as radiotherapy, which can leverage the findings of this study, can help clinical evaluation of the risks and benefits of liver intervention.

이진 분할(binary segmentation)은 특히 작은 대상 영역을 놓치면 신뢰성을 크게 훼손할 수 있는 응용 분야에서 거짓 음성(False Negatives, FNs) 문제를 지속적으로 겪고 있다. 특히 희소한 토지 피복을 탐지하는 작업에서는 FN이 실제 대상을 모델이 검출하지 못한 것으로, 모델 성능의 신뢰성을 저해하는 핵심적인 요소로 간주된다. 이에 본 연구에서는 1단계 학습(Stage 1)에서 예측된 FP 및 FN 영역을 자동으로 축적하고, 이를 보조 감독 신호(auxiliary supervisory signals)로 활용하는 두 번째 출력 헤드를 통해 2단계 학습(Stage 2)에서 재학습하도록 하는 2단계 U-Net 학습 전략을 제안한다. 이 접근법은 모델 구조를 변경하지 않고도 불확실하거나 모호한 영역에 대한 학습 반영을 강화할 수 있다는 점에서 실용적이다. 제안된 방법은 대한민국의 고해상도 항공영상을 활용한 희소 분포 삼계(인삼) 재배지 탐지에 적용되었으며, FN을 감소시키고 FP의 소폭 증가를 허용하는 대신 평균 교차합집합(Intersection over Union, IoU)을 약 0.10 개선하는 결과를 보였다. 본 전략은 학습 방식만을 수정하는 방법으로, 다양한 U-Net 변형 모델과 호환 가능하며 FN 억제와 예측 안정성이 중요한 다른 도메인에도 일반화할 수 있다. 이러한 결과는 누적된 오차 영역으로부터의 학습이 복잡한 실제 환경에서 보다 신뢰할 수 있는 분할 성능을 달성하는 효율적이고 실용적인 경로가 될 수 있음을 보여준다.

Binary segmentation faces persistent challenges from false negatives (FNs), particularly in applications where missing small target regions can undermine reliability. FNs are critical in sparse land-cover mapping tasks because they represent true targets that models fail to detect. Here, we propose a two-stage U-Net training strategy that automatically accumulates FP and FN regions from Stage 1 predictions and reuses them as auxiliary supervisory signals through a second output head during Stage 2 learning. This enhances the model's ability to revisit uncertain or ambiguous areas without any architectural modification. The method was applied to mapping sparsely distributed ginseng fields from high-resolution aerial imagery from South Korea. The approach substantially reduced FNs, tolerated a modest increase in FPs, and improved Intersection over Union (IoU) by approximately 0.10 across various training conditions. Because the strategy only modifies the training scheme and not the model architecture, it is compatible with U-Net variants and generalizable to other domains where FN suppression and stability are essential. The findings demonstrate that learning from accumulated error regions offers an efficient and practical pathway toward more reliable segmentation in complex real world environments.

최근 딥러닝 기술이 급속히 발전하며 현대 사회의 다양한 응용분야에서 빠르게 적용되고 있다. 특히 영상 기반의 딥러닝 기술은 자연어 처리와 함께 인공지능 기술의 핵심 연구 분야로 많은 연구가 진행되고 있다. 논문에서는 최근 많은 연구가 진행되고 있는 영상의 의미적 분할 (Semantic Segmentation) 성능을 향상하기 위한 연구를 진행한다. 특히 모델에서 고정밀의 의미적 분할을 수행할 수 있도록 추가적인 정보로써 의사 깊이맵 (Pseudo Depth-Map)을 활용하는 방법을 제안하였다. 더불어, 의사 깊이맵을 모델 상에서 효과적으로 학습시키기 위하여 다중 디코더 모델과 학습 효율을 높이는 학습 스케줄링 전략을 제안한다. 의사 깊이맵과 다중 디코더 모델 기반의 제안 모델은 기존 의미적 분할 모델과 비교하여 iIoU 기준 2%의 성능 향상을 보였다.

최근 들어 위성의 소형화 및 군집화로 인해 고해상도 위성영상의 수요가 증가함에 따라 사용자가 보다 쉽게 위성영상을 활용할 수 있도록 지원하기 위한 다양한 노력들이 이루어지고 있다. 이에 한국항공우주연구원에서는 공공분야 사용자의 편의성을 향상시키고 위성영상 활용을 활성화하기 위하여 매년 한반도 모자이크 영상을 제작하여 제공하고 있다. 이러한 한반도 모자이크 영상의 활용도를 높이기 위하여 모자이크 영상을 활용한 위성영상 분할 및 분류 연구를 수행하고자 하였다. 그러나 모자이크 영상은 R, G, B 밴드만 보유하고 있고, 영상 융합 및 컬러 밸런싱 등의 과정을 적용하였기 때문에 분광 정보에 왜곡이 존재한다는 한계점이 있어 이를 보완하고자 R, G, B 밴드를 활용하여 다양한 지수를 추출하여 영상 분할 및 분류에 적용하였다. 연구 결과, 모자이크 영상만을 활용한 영상 분류 결과의 정확도는 약 72%인 반면, R, G, B 밴드에서 추출한 지수를 함께 활용한 영상 분류 결과의 정확도는 약 79%로 나타났다. 이를 통해 한반도 모자이크 영상을 활용하여 영상 분류를 수행할 때 R, G, B 밴드에서 추출한 지수를 함께 활용할 경우 영상 분류 결과를 향상시킬 수 있다는 것을 확인할 수 있었다. 이러한 연구 결과는 모자이크 영상뿐만 아니라 분광 정보가 왜곡되거나 R, G, B 밴드만을 제공하는 영상에도 적용될 수 있을 것으로 기대된다.

In recent years, as the demand of high-resolution satellite images increases due to the miniaturization and constellation of satellites, various efforts to support users to utilize satellite images more conveniently are performed. Accordingly, the Korea Aerospace Research Institute produces and provides mosaic images on the Korean Peninsula every year to improve the convenience of users in the public sector and activate the use of satellite images. In order to increase the utilization of mosaic images on the Korean Peninsula, a study on satellite image segmentation and classification using mosaic images was attempted. However, since mosaic images provide only R, G, and B bands and processes such as image sharpening and color balancing are applied, there is a limitation that the spectral information of original images is distorted, so various indices were extracted and classified using R, G, and B bands to compensate for this. As a result of the study, the accuracy of image classification results using only mosaic images was about 72%, while the accuracy of image classification results using indices extracted from R, G, and B bands together was about 79%. Through this, it was confirmed that when performing image classification using mosaic images on the Korean Peninsula, the image classification results can be improved if the indices extracted from R, G, and B bands are used together. These research results are expected to be applied not only to mosaic images but also to images in which spectral information is limited or only R, G, and B bands are provided.

목 적: 인공지능 (Artificial intelligence, AI)의 발전과 함께, 방사선종양학과의 임상 영역에서는 AI 딥 러닝 (Deep learning, DL) 기반의 자동 세분화 윤곽선 생성 시스템(Auto-segmentation Contouring System, ASCS)이 널리 상용되고 있다. 이에 발맞추어, 본 연구는 미국 방사선종양학임상연구회 가이드라인(RTOG)에 따라 방사선종양학과 전문의에 의해 수동으로 얻은 방사선 치료 부위의 위험 장기 (Organ at risk, OAR) 수동 윤곽선과 ASCS에 의해 생성된 OAR 윤곽선을 비교함으로서 AI DL 기반의 ASCS의 정확도를 평가하고자 한다. 대상 및 방법: 2022년 1월부터 2024년 1월까지 유방암을 진단받고, 방사선 치료를 받은 환자 중 임상의 1명에 의해 수동으로 생성된 OAR 윤곽선이 확인되는 109명의 환자를 추출하였다. 이 환자들의 Planning CT에 본원에서 사용 중인 ASCS인 OncoStudio 2.1.1Ver를 통하여 OAR 자동 윤곽선 생성(Auto-contouring)을 진행하였다. 이어서 이미 그려진 임상의에 의한 수동 OAR 중 치료 동측 폐(Lung)와 심장(Heart)에 대한 윤곽선에 대해 ASCS를 통해 얻은 폐와 심장의 윤곽선의 정확도를 평가하기 위해 다이스 유사 계수(Dice Similarity Coefficient, DSC), 평균 표면 거리(Mean Surface Distance, MSD), 95% 하우스도르프 거리(Hausdorff distance, HD) 값을 도출하였다. 결 과: 수동으로 기생성된 OAR 윤곽선에 대해 ASCS를 통해 생성한 OAR 윤곽선의 DSC, MSD, 그리고 95%HD의 평가 지표값은 폐 윤곽선에서 0.972, 0.922, 그리고 4.801 각각 보고되었고, 심장 윤곽선에서 0.919, 2.329, 그리고 7.027 각각 보고되었다. DSC, MSD, 그리고 95%HD의 세 평가 지표 모두에서 수동으로 그려진 OAR 윤곽선에 대한 ASCS 기반의 OAR 윤곽선의 정확도가 폐 윤곽선에 비해 심장 윤곽선에서 떨어지는 것으로 나타났다. 결 론: ASCS를 이용하여 생성된 OAR 윤곽선의 정확성을 DSC, MSD, 그리고 95%HD의 값을 바탕으로 분석하였을 때 높은 것을 확인하였다. 다만, 폐 윤곽선에서는 높은 정확도를 보여주었으나, 심장 윤곽선의 경우 정확도가 상대적으로 다소 떨어지는 것을 확인하였다. 이에 ASCS 기반의 자동 윤곽선 생성은 OAR 윤곽선 설정에 도움을 주지만, OAR에 따라 윤곽선을 확인하고 수정하는 데 소요되는 노력과 시간이 다를 것으로 판단된다.

Purpose: With the advancement of artificial intelligence (AI), deep learning (DL)-based auto-segmentation contouring systems (ASCS) have been widely implemented in clinical radiation oncology. This study aimed to evaluate the accuracy of an AI DL-based ASCS by comparing organ-at-risk (OAR) contours automatically generated by the ASCS with manually delineated OAR contours drawn by a radiation oncologist, according to the Radiation Therapy Oncology Group (RTOG) guidelines. Materials & Methods: Between January 2022 and January 2024, a total of 109 patients diagnosed with breast cancer and treated with radiotherapy were retrospectively selected. All patients had manually delineated OAR contours generated by a single clinician. Auto-contouring of OARs was performed on each patient's planning CT using the institution's ASCS (OncoStudio version 2.1.1). The accuracy of the automatically generated contours for the ipsilateral lung and heart was quantitatively evaluated against the manual contours using the Dice Similarity Coefficient (DSC), Mean Surface Distance (MSD), and 95% Hausdorff Distance (HD). Result: For the manually delineated OAR contours, the automatically generated contours by the ASCS showed DSC, MSD, and 95% HD values of 0.972, 0.922, and 4.801, respectively, for the lung, and 0.919, 2.329, and 7.027, respectively, for the heart. All three evaluation metrics indicated that the accuracy of ASCS-based contours was lower for the heart than for the lung. Conclusion: Based on DSC, MSD, and 95% HD analyses, the ASCS demonstrated high overall accuracy in automatic OAR contour generation. However, while the system showed excellent performance in lung contouring, the accuracy for the heart was relatively lower. Therefore, although ASCS can assist in OAR delineation, the level of manual verification and correction required may vary depending on the specific organ.

Since mobile phones are used as common communication devices, their applications are increasingly important to human's life. Using smart-phones camera to collect daily life environment's information is one of targets for many applications such as text recognition, object recognition or context awareness. Studies have been conducted to provide important information through the recognition of texts, which are artificially or naturally included in images and movies acquired from mobile phones. In this study, a character segmentation method that improves character-recognition accuracy in images obtained from mobile phone cameras is proposed. The proposed method first classifies texts in a given image to printed letters and handwritten letters since segmentation approaches for them are different. For printed letters, rough segmentation process is conducted, then the segmented regions are integrated, deleted, and re-segmented. Segmentation for the handwritten letters is performed after skews are corrected and the characters are classified by integrating them. The experimental result shows our method achieves a successful performance for both printed and handwritten letters as 95.9% and 84.7%, respectively.

본 논문에서는 대상지역에 대한 영상을 다양한 가중치의 조합의 경우를 고려하여 객체 단위로 분할하게 되며 분할된 객체에 대하여 상호관계를 분석하여 수치적으로 표현하였다. 또한 최종적인 객체 기반영상분류에서 높은 정확도를 확보할 수 있는 가중치의 조합을 산정하였다. 연구에 사용된 영상은 Landsat-7/ETM 영상으로 대상 지역의 면적은 $11{\times}14$ Km이며 밴드 2, 3, 4의 조합을 사용하였다. 객체 간 계산은 Moran's I와 객체 내부 분산(Intrasegment Variance)을 이용하였다. 대상지역에 대하여 총 75개의 가중치 조합을 사용하여 75개의 객체 분할 영상을 생성하였다. 객체 분할 영상 중에 최종적인 영상 분류 시 높은 정확도가 예상되는 가중치 조합, 중간 정도 정확도가 예상되는 가중치 조합 그리고 낮은 정도 정확도가 예상되는 가중치 조합을 7개 선택하여 최종적인 객체기반 영상분류를 시행하고 그 정확도를 비교하였다. 정확도의 비교 결과, 가장 높은 정확도가 예상되는 가중치 조합의 객체 분할 영상의 경우 객체 기반 영상 분류 시 85% 이상의 정확도를 나타내었으며 반대로 낮은 경우는 분류 시 50% 정도의 분류 정확도를 나타내었다.

The overall objective of this research was to investigate various combination of segmentation parameters and to improve classification accuracy of object-oriented classification. This research presents a method for evaluation of segmentation parameters by calculating Moran's I and Intrasegment Variance. This research used Landsat-7/ETM image of $11{\times}14$ Km developed area in Ansung, Korea. Segmented images are generated by 75 combinations of parameter. Selecting 7 combinations of high, middle and low grade expected classification accuracy was based on calculated Moran's I and Intrasegment Variance. Selected segmentation images are classified 4 classes and analyzed classification accuracy according to method of objected-oriented classification. The research result proved that classification accuracy is related to segmentation parameters. The case of high grade of expected classification accuracy showed more than 85% overall accuracy. On the other hand, low ado showed around 50% overall accuracy.

본 연구는 딥러닝 모델(deep learning model)을 활용하여 토지피복분류를 수행하였으며 입력 이미지의 크기, Stride 적용 등 데이터세트(dataset)의 조절을 통해 토지피복분류를 위한 최적의 딥러닝 모델 선정을 목적으로 하였다. 적용한 딥러닝 모델은 3종류로 Encoder-Decoder 구조를 가진 U-net과 DeeplabV3+, 두 가지 모델을 결합한 앙상블(Ensemble) 모델을 활용하였다. 데이터세트는 RapidEye 위성영상을 입력영상으로, 라벨(label) 이미지는 Intergovernmental Panel on Climate Change 토지이용의 6가지 범주에 따라 구축한 Raster 이미지를 참값으로 활용하였다. 딥러닝 모델의 정확도 향상을 위해 데이터세트의 질적 향상 문제에 대해 주목하였으며 딥러닝 모델(U-net, DeeplabV3+, Ensemble), 입력 이미지 크기(64 × 64 pixel, 256 × 256 pixel), Stride 적용(50%, 100%) 조합을 통해 12가지 토지피복도를 구축하였다. 라벨 이미지와 딥러닝 모델 기반의 토지피복도의 정합성 평가결과, U-net과 DeeplabV3+ 모델의 전체 정확도는 각각 최대 약 87.9%와 89.8%, kappa 계수는 모두 약 72% 이상으로 높은 정확도를 보였으며, 64 × 64 pixel 크기의 데이터세트를 활용한 U-net 모델의 정확도가 가장 높았다. 또한 딥러닝 모델에 앙상블 및 Stride를 적용한 결과, 최대 약 3% 정확도가 상승하였으며 Semantic Segmentation 기반 딥러닝 모델의 단점인 경계간의 불일치가 개선됨을 확인하였다.

The purpose of this study was to construct land cover maps using a deep learning model and to select the optimal deep learning model for land cover classification by adjusting the dataset such as input image size and Stride application. Two types of deep learning models, the U-net model and the DeeplabV3+ model with an Encoder-Decoder network, were utilized. Also, the combination of the two deep learning models, which is an Ensemble model, was used in this study. The dataset utilized RapidEye satellite images as input images and the label images used Raster images based on the six categories of the land use of Intergovernmental Panel on Climate Change as true value. This study focused on the problem of the quality improvement of the dataset to enhance the accuracy of deep learning model and constructed twelve land cover maps using the combination of three deep learning models (U-net, DeeplabV3+, and Ensemble), two input image sizes (64 × 64 pixel and 256 × 256 pixel), and two Stride application rates (50% and 100%). The evaluation of the accuracy of the label images and the deep learning-based land cover maps showed that the U-net and DeeplabV3+ models had high accuracy, with overall accuracy values of approximately 87.9% and 89.8%, and kappa coefficients of over 72%. In addition, applying the Ensemble and Stride to the deep learning models resulted in a maximum increase of approximately 3% in accuracy and an improvement in the issue of boundary inconsistency, which is a problem associated with Semantic Segmentation based deep learning models.

본 연구의 목적은 50 cm급의 worldview-1 입체영상을 스테레오 매칭 알고리듬을 이용하여 DSM을 생성하고 정확도를 확인하고자 한다. DSM을 생성하기 위하여, 먼저, RPC 블록모델링을 수행하여 RPC 오차를 보정하였으며, 에피폴라 영상 생성 후 스테레오 매칭 알고리듬인 SGM을 이용하여 영상 매칭을 수행하였다. SGM에 사용된 COST는 CENSUS를 사용하였으며, COST 누적을 위하여 4방향과 8방향을 적용하였다. 생성된 DSM의 품질과 정확도를 확인하기 위하여 LiDAR DSM과 상용 SW로 생성한 DSM을 비교하였다. 실험 결과 지역에 따라 생성된 DSM의 수직정확도는 4방향의 COST 적용 시 RMSE 1.647 m ~ 3.689 m로 나타났으며, 8방향의 COST 적용 시 RMSE 1.550 m ~ 3.106 m로 나타났다.

The purpose of this study is to generate DSM using the stereo matching algorithm of worldview-1 stereo images and verify the accuracy of the generated DSM. To generate DSM, RPC block modeling was performed to correct RPC errors, and image matching was performed using SGM, which is a stereo matching algorithm after the epipolar image was generated. The COST for SGM was calculated by using CENSUS, and 4-paths and 8-paths were applied for COST aggregation in SGM. To verify the quality and accuracy of the generated DSM, it was compared with the LiDAR-derived DSM and the DSM generated by commercial SW. The results showed that the vertical accuracy of the generated DSM using 4-paths of COST aggregation was 1.647 m to 3.689 m (RMSE). In case of using 8-paths of COST aggregation was 1.550 m to 3.106 m (RMSE).

본 연구에서는 Attention Mechanism이 적용된 UNet 모델을 활용하여 유방 초음파 영상에서 병변 분할 성능을 분석하였다. 기존 연구에서는 IoU를 주요 성능 지표로 활용하여 모델의 성능을 평가하였으나, 본 연구에서는 기존 연구의 IoU뿐만 아니라 Dice Coefficient, Accuracy, F1 Score를 기반으로 한 개별 모델을 구축하여 성능을 종합적으로 비교하였다. 실험 결과, IoU 및 Dice Coefficient 기반 모델이 가장 높은 분할 성능을 보였으며, Accuracy 기반 모델은 픽셀 단위의 정확도는 높았지만 병변 경계를 명확하게 구별하는 능력이 부족하였다. F1 Score 기반 모델은 높은 Recall을 기록하여 False Negative를 줄이는 데 효과적이었으나, Precision이 낮아 False Positive 발생 가능성이 증가하는 경향을 보였다. 본 연구를 통해 Attention Mechanism이 적용된 UNet 모델이 기존 IoU 기반 방법보다 병변 분할 성능에서 향상된 결과를 보이며, IoU 및 Dice Coefficient 기반 학습이 병변 분할 성능을 극대화하는 데 효과적임을 확인하였다.

In this study, we analyzed the lesion segmentation performance in breast ultrasound images using a UNet model enhanced with an Attention Mechanism. Previous studies primarily used IoU as the key performance metric to evaluate model performance; however, in this study, we constructed individual models based not only on IoU from previous studies but also on Dice Coefficient, Accuracy, and F1 Score to comprehensively compare their performance. The experimental results showed that models based on IoU and Dice Coefficient achieved the highest segmentation performance, while the Accuracy-based model exhibited high pixel-wise accuracy but lacked the ability to clearly distinguish lesion boundaries. The F1 Score-based model demonstrated higher Recall, making it effective in reducing False Negatives, but its lower Precision led to an increased likelihood of False Positives. Through this study, we confirmed that the UNet model enhanced with an Attention Mechanism outperforms conventional IoU-based methods in lesion segmentation performance. Furthermore, IoU and Dice Coefficient-based training were found to be the most effective in maximizing lesion segmentation performance.