Earticle

Computed Tomography (CT) 영상에서 Gaussian noise는 영상이 획득되는 과정에서 불가피하게 발생되며, 이는 segmentation의 성능 및 진단의 정확도를 저하시키는 요인으로 작용한다. 따라서, 본 연구에서는 denoising filter의 종류에 따른 영역확장 (region growing, RG) 기반 segmentation 알고리즘의 성능을 정량적으로 비교 평가하였다. 실험을 위해 Shepp-Logan phantom 영상으로부터 0.05의 표준편차 값을 갖는 Gaussian noise를 부가하고 각각 mean, median, Wiener filter를 적용하였다. 이후, RG 기반 segmentation 알고리즘을 적용하였으며, 이에 대한 정량적 평가를 위해 F1-score와 IoU를 측정하였다. 결과적으로, F1-score 및 IoU 인자 모두 noisy, mean filter, median filter 그리고 Wiener filter 순으로 우수한 값을 보였다. 특히, Wiener filter를 적용하였을 때 noisy 영상과 비교하여 F1-score 및 IoU 인자는 각각 약 1.42배 및 1.75배 향상됨을 확인하였다. 결론적으로, CT 영상의 segementation 정확도를 향상시키기 위해 적합한 denoising filter를 적용해야 함을 증명하였다.

Segmentation is very important in accurate surgical planning such as liver transplantation and resection and identification of lung nodules. In a computed tomography image, Gaussian noise inevitably occurs in the process of acquiring the image. Gaussian noise can interfere with segmentation and interfere with accurate diagnosis, resulting in fatal consequences for patients. Therefore, this study attempted to compare and evaluate which filter is most effective for denoising in performing segmentation. For the study, Gaussian noise was added to Shepp-Logan phantom and mean, median, and Wiener filter were applied, respectively. Thereafter, segmentation was performed using the thresholding technique to evaluate using F1-score and IoU to find the most suitable filter for denoising. As a results, the F1-score was 0.62 for noise-added image, 0.697 for mean filter, 0.858 for median filter, and 0.88 for Wiener filter. The IoU was 0.449 for noise-added image, 0.535 for mean filter, 0.752 for median filter, and 0.785 for Wiener filter. Therefore, the Wiener filter was confirmed to be the most effective for denoising, and it was confirmed that the F1-score improved 1.42 times and the IoU 1.75 times compared to the noise-added image. In conclusion, various denoising filter methods were analyzed to improve the accuracy of segmentation, and Wiener filter is expected to be used in an efficient way.

체성분 분석은 다양한 질병의 진단, 예후 예측 및 치료 반응 평가에 중요한 역할을 한다. 특히 L3 요추 단면에서 피하지방, 내장지방, 골격근의 정량적 평가는 비만, 근감소증, 심혈관 질환 등과 밀접하게 연관되어 있다. 본 연구에 서는 VerSe 공개 데이터셋에 TotalSegmentator를 활용하여 자동 분할을 진행하였다. 자동 분할 후, ITK-SNAP 을 활용하여 과분할 및 누락 영역을 보정하였다. 예측 모델의 성능을 평가하기 위해 Dice 유사도 점수, 민감도, 특 이도 비교 및 Bland-Altman 분석을 수행하였다. 보정한 데이터로 학습한 모델이 보정하지 않은 데이터로 학습한 모델보다 Dice 유사도 점수 측면에서 피하지방, 골격근 부위에서는 높은 성능을 보였고, 조직이 아닌 부위의 분할 오류가 줄었다. 경량화 U-Net의 경우 기본 U-Net보다 Dice 유사도 점수와 특이도가 높았고, Bland-Altman 분 석에서 내장지방의 편향과 일치도 한계 폭이 줄었다. 결론적으로 TotalSegmentator로 자동 분할한 결과를 보정하 지 않고 사용해도 피하지방, 내장지방, 골격근 부위의 Dice 유사도 점수가 0.8이상이 나오지만, 성능을 더 높이고 관심 조직이 아닌 부위의 분할 오류를 줄이려면 보정한 데이터로 학습한 모델을 사용하는 것을 권장한다. 경량화 모 델도 분할 성능이 기본 모델 못지 않으므로 컴퓨팅 자원이 제한된 기기에서 사용을 권장한다.

Body composition analysis plays an important role in diagnosis, prognosis, and treatment monitoring of various diseases. In particular, quantitative evaluation of subcutaneous fat, visceral fat, and skeletal muscle in the third lumbar vertebral cross section is closely related to obesity, sarcopenia, and cardiovascular disease. In this study, the VerSe computed tomography(CT) spine dataset was used, and automatic segmentation was performed using TotalSegmentator. After automatic segmentation, over- and under-segmented regions were manually corrected by using ITK-SNAP. For evaluation of the performance of prediction models, we performed comparisons of Dice similarity score, sensitivity, and specificity as well as Bland-Altman analysis. The model trained on corrected data achieved higher Dice similarity scores in the subcutaneous fat and skeletal muscle than the model trained on uncorrected data, and it resulted in reduction in segmentation errors outside the regions of interest. The lightweight U-Net produced higher Dice similarity scores than the standard U-Net, and it resulted in reductions of the bias and limits of agreement in the visceral fat in Bland- Altman analysis. In conclusion, the use of TotalSegmentator without correction of segmentation errors still produced Dice similarity scores higher than 0.8 in all three regions of interest, but if one wants to improve segmentation performance and reduce segmentation errors outside the regions of interest, we recommend training the models with corrected segmentation masks. We also recommend using lightweight U-Net models, which are on par with standard U-Net models, in the device with limited computing resources.

도로 위 화물차의 적재 불량은 심각한 교통사고를 유발할 수 있기 때문에, 이를 예방하고 안전성을 높이기 위해서는 자동화된 적재 불량 단속 시스템이 필요하다. 이러한 배경에서 본 연구는 적재 불량 중 특히 적재 크기 초과 유형을 판단하는 사전 학습된 EfficientNet 기반 두 가지 딥러닝 모델을 제시하였다. 첫 번째 모델은 화물차의 영역을 bounding box로 추출한 화물 차 이미지를 입력으로 사용하였으며, 두 번째 모델은 bounding box 기반 추출 화물차 이미지와 이미지 분할(image segmentation) 기반 화물차 적재 영역 이미지를 함께 사용하였다. 평가 결과, 두 모델 모두 정확도 0.95 이상의 높은 분류 성능을 보여주었으나, 두 번째 모델이 전반적으로 더 높은 성능을 보였다. 즉, 화물차 이미지와 함께 이미지 분할을 통해 적재함 이미지를 함께 제공할 경우 화물차 적재 크기 초과 여부를 판단하는 성능이 높아지는 것을 확인하였다.

Improper loading of cargo trucks can lead to serious traffic accidents. To prevent such accidents and improve safety, an automated system for detecting overloaded trucks is essential. This study proposes two deep learning models based on a pre-trained EfficientNet to detect cargo that exceeds size limits. The first model uses cargo truck images extracted with a bounding box around the truck's area, whereas the second model utilizes both bounding box and a loading area image extracted via image segmentation. Evaluation results show that both models demonstrated high performance with classification accuracy over 0.95; however, the second model showed overall superior performance. Specifically, it confirmed that incorporating image segmentation from the loading area along with the truck image improved performance when determining whether cargo sizes exceeded their limits.

본 논문에서는 효율적인 차량 객체를 추적하는 특징점 매칭을 이용한 다중 차량 객체 검 출 알고리즘을 제안한다. 제안하는 알고리즘은 효율적인 차량 객체 추적을 위해 FAST 알고 리즘을 이용해서 차량의 특징점을 추출한다. 그리고 5X5 영역으로 분할 된 영상에서 특징점 이 포함되면 True 포함되지 않으면 False로 해당 영역을 검은색으로 후처리하여 차량 객체을 제외한 불필요한 객체 정보를 제거한다. 그리고 후처리 된 영역을 차량의 최대 탐색창 크기 로 설정하고, 차량의 최외각 특징점을 이용한 최소 탐색창을 설정하여 Mean-Shift 알고리즘 의 탐색창 크기에 대한 단점을 보완하여 차량 객체 추적을 한다. 제안한 방법의 성능 평가하 기위해 SIFT, SURF 알고리즘을 비교하여 실험한다. 그 결과 SIFT 알고리즘에 비해서 약 4배 빠르고 SUFR 알고리즘의 처리 과정 보다는 효율적으로 검출하는 장점이 있다

In this paper, we propose a multi-vehicle object detection algorithm using feature point matching that tracks efficient vehicle objects. The proposed algorithm extracts the feature points of the vehicle using the FAST algorithm for efficient vehicle object tracking. And True if the feature points are included in the image segmented into the 5X5 region. If the feature point is not included, it is processed as False and the corresponding area is blacked to remove unnecessary object information excluding the vehicle object. Then, the post processed area is set as the maximum search window size of the vehicle. And A minimum search window using the outermost feature points of the vehicle is set. By using the set search window, we compensate the disadvantages of the search window size of mean-shift algorithm and track vehicle object. In order to evaluate the performance of the proposed method, SIFT and SURF algorithms are compared and tested. The result is about four times faster than the SIFT algorithm. And it has the advantage of detecting more efficiently than the process of SUFR algorithm.

영상 분할은 의료 영상 처리에서 중요한 역할을 수행하지만 영상 분할 알고리즘의 많은 연 산량은 실생활에서의 사용 을 제약하는 요인이 된다. 따라서 본 논문에서는 본 논문에서는 고해상도 의료 영상 분할 알 고리즘의 실시간 처리를 위해 GPGPU(general purpose graphics processing unit)상에서 영상 분할 알고리즘인 Fuzzy C-Means (FCM)의 병렬구현 방법을 제안한다. 본 논문에서는 GPU상에서 FCM 알고리즘의 고속화를 위해 대상 영상을 재 구성하고, 단일 쉐이더 코어에서 필요한 레지스터 개수를 바탕으로 FCM 알고리즘을 위한 최적의 쓰레드(thread) 와 블록(block)의 크기를 계산한다. 성능 평가를 위해 본 논문에서는 총 5가지 크기의 의료 영상(64×64, 128×128, 256×256, 512×512, 1,024×1,024)을 이용하여 단일 CPU대비 향상된 실행 시간을 측정하였고, 모 의실험 결과 64x64 영상을 제외한 나머지 영상에서 단일 CPU 보다 약 4배 높은 실행 시간 의 향상을 나타내었다.

Image segmentation plays a crucial role in numerous biomedical imaging applications. However, its high computational complexities require substantial amount of time and have limited its use in real applications. With this reason, this paper implements a fuzzy c-means (FCM) clustering algorithm on general purpose graphics processing unit (GPGPU) in order to run it in real time. This paper first reconfigures a target medical image for real-time processing of FCM, and calculates optimal sizes of threads and blocks for FCM based on the number of registers required by a shader core. To evaluate the proposed approach on GPGPU, this paper measures execution times to complete FCM on GPGPU by utilizing medical images at different resolutions (64×64, 128×128, 256×256, 512×512, 1,024×1,024). Experimental results indicate that execution times on GPGPU are approximately 4-fold faster than those on a single CPU for all of medical images except an 64×64 medical image.

Although huge progress has been made in current image segmentation work, there are still no efficient segmentation strategies for tree image which is taken from natural environment and contains complex background. To improve those problems, we propose a method for tree image segmentation combining adaptive mean shifting with image abstraction. Our approach perform better than others because it focuses mainly on the background of image and characteristics of the tree itself. First, we abstract the original tree image using bilateral filtering and image pyramid from multiple perspectives, which can reduce the influence of the background and tree canopy gaps on clustering. Spatial location and gray scale features are obtained by step detection and the insertion rule method, respectively. Bandwidths calculated by spatial location and gray scale features are then used to determine the size of the Gaussian kernel function and in the mean shift clustering. Furthermore, the flood fill method is employed to fill the results of clustering and highlight the region of interest. To prove the effectiveness of tree image abstractions on image clustering, we compared different abstraction levels and achieved the optimal clustering results. For our algorithm, the average segmentation accuracy (SA), over-segmentation rate (OR), and under-segmentation rate (UR) of the crown are 91.21%, 3.54%, and 9.85%, respectively. The average values of the trunk are 92.78%, 8.16%, and 7.93%, respectively. Comparing the results of our method experimentally with other popular tree image segmentation methods, our segmentation method get rid of human interaction and shows higher SA. Meanwhile, this work shows a promising application prospect on visual reconstruction and factors measurement of tree.

With the development of artificial intelligence technology, various methods have been developed for recognizing objects in images using machine learning. Image segmentation is the most effective among these methods for recognizing objects within an image. Conventionally, image datasets of various classes are trained simultaneously. In situations where several classes require segmentation, all datasets have to be trained thoroughly. Such repeated training results in low training efficiency because most of the classes have already been trained. In addition, the number of classes that appear in the datasets affects training. Some classes appear in datasets in remarkably smaller numbers than others, and hence, the training errors will not be properly reflected when all the classes are trained simultaneously. Therefore, a new method that separates some classes from the dataset is proposed to improve efficiency during training. In addition, the accuracies of the conventional and proposed methods are compared.

Three-dimensional (3D) medical image segmentation is vital in clinical applications, but faces challenges like computational complexity and spatial context loss in 2D processing. To address these issues, we propose 3DAS2D, which processes 3D images as 2D slices while retaining volumetric data for better accuracy. Our approach integrates modules like Neighbour Attention, Memory Attention, Memory Bank, and Mask Encoder to capture inter-slice dependencies and historical context within a 2D prediction framework. Evaluated on four datasets covering prostate, cardiac, and lung tumour tasks, 3DAS2D performed comparably to 3D methods. This work offers an efficient, adaptable solution bridging 2D efficiency and 3D accuracy, with significant clinical potential.

Explainable AI (XAI) has found numerous applications in computer vision. While image classification-based explainability techniques have garnered significant attention, their counterparts in semantic segmentation have been relatively neglected. Given the prevalent use of image segmentation, ranging from medical to industrial deployments, these techniques warrant a systematic look. In this paper, we present the first comprehensive survey on XAI in semantic image segmentation. We analyze and categorize the literature based on application categories and domains, as well as the evaluation metrics and datasets used. We also propose a taxonomy for interpretable semantic segmentation, and discuss potential challenges and future research directions.

In order to improve the edge segmentation effect of the level set image segmentation and avoid the influence of the initial contour on the level set method, a saliency level set image segmentation model based on local Renyi entropy is proposed. Firstly, the saliency map of the original image is extracted by using saliency detection algorithm. And the outline of the saliency map can be used to initialize the level set. Secondly, the local energy and edge energy of the image are obtained by using local Renyi entropy and Canny operator respectively. At the same time, new adaptive weight coefficient and boundary indication function are constructed. Finally, the local binary fitting energy model (LBF) as an external energy term is introduced. In this paper, the contrast experiments are implemented in different image database. The robustness of the proposed model for segmentation of images with intensity inhomogeneity and complicated edges is verified.

Level set methods are the numerical techniques for tracking interfaces and shapes. They have been successfully used in image segmentation. A new variational level set evolving algorithm without re-initialization is presented in this paper. It consists of an internal energy term that penalizes deviations of the level set function from a signed distance function, and an external energy term that drives the motion of the zero level set toward the desired image feature. This algorithm can be easily implemented using a simple finite difference scheme. Meanwhile, not only can the initial contour can be shown anywhere in the image, but the interior contours can also be automatically detected.

An image semantic segmentation model is proposed based on improved ENet network in order to achieve the low accuracy of image semantic segmentation in complex environment. Firstly, this paper performs pruning and convolution optimization operations on the ENet network. That is, the network structure is reasonably adjusted for better results in image segmentation by reducing the convolution operation in the decoder and proposing the bottleneck convolution structure. Squeeze-and-excitation (SE) module is then integrated into the optimized ENet network. Small-scale targets see improvement in segmentation accuracy via automatic learning of the importance of each feature channel. Finally, the experiment was verified on the public dataset. This method outperforms the existing comparison methods in mean pixel accuracy (MPA) and mean intersection over union (MIOU) values. And in a short running time, the accuracy of the segmentation and the efficiency of the operation are guaranteed.

This paper is an attempt to design segmentation method based on fully convolutional networks (FCN) and attention mechanism. The first five layers of the Visual Geometry Group (VGG) 16 network serve as the coding part in the semantic segmentation network structure with the convolutional layer used to replace pooling to reduce loss of image feature extraction information. The up-sampling and deconvolution unit of the FCN is then used as the decoding part in the semantic segmentation network. In the deconvolution process, the skip structure is used to fuse different levels of information and the attention mechanism is incorporated to reduce accuracy loss. Finally, the segmentation results are obtained through pixel layer classification. The results show that our method outperforms the comparison methods in mean pixel accuracy (MPA) and mean intersection over union (MIOU).

This work presents an efficient LTP-based sharpness measure for blur detection and segmentation. The proposed method transforms each pixel into ternary codes depending on the differences of intensity of the central pixel with the neighborhood pixels. These ternary codes have been converted into lower and upper binary patterns. Among these, the non-uniform patterns have been exploited to compute the blur measure and blur segmentation. The proposed methodology performs segmentation without having any explicit information about the type and level of the blur. Experimental results reveal that the proposed method outperforms the state-of-the-art blur detection and segmentation methods.

A variety of medical service applications in the field of the Internet of Things (IoT) are being studied. Segmentation is important to identify meaningful regions in images and is also required in 3D images. Previous methods have been based on gray value and shape. The Visible Korean dataset consists of serially sectioned high-resolution color images. Unlike computed tomography or magnetic resonance images, automatic segmentation of color images is difficult because detecting an object's boundaries in colored images is very difficult compared to grayscale images. Therefore, skilled anatomists usually segment color images manually or semi-automatically. We present an out-of-core 3D segmentation method for large-scale image datasets. Our method can segment significant regions in the coronal and sagittal planes, as well as the axial plane, to produce a 3D image. Our system verifies the result interactively with a multi-planar reconstruction view and a 3D view. Our system can be used to train unskilled anatomists and medical students. It is also possible for a skilled anatomist to segment an image remotely since it is difficult to transfer such large amounts of data.

본 논문은 지문 데이터베이스를 평가하는데 가장 큰 영향을 미치는 image quality를 측정하는 새로운 방법을 제안한다. 본 논문에서는 image quality를 측정하는 hybrid segmentation 방법을 소개하고, 다양한 지문 데이터베이스에 대해 실험한 결과를 분석한다. 개발한 방법의 객관적인 평가를 위해 NIST에서 제공하는 NFIQ 프로그램을 통해 얻은 결과와 variance와 coherence의 fusion을 이용한 hybrid segmentation 결과를 비교한다. NFIQ는 지문 영상의 품질을 정확하게 측정하지만 결과가 $1{\sim}5$로 세분화되어 있지 못한 문제점을 가지고 있다. 반면 제안하는 hybrid 방법은 NFIQ보다 더 정확하고 세분화된 평가 결과를 제공한다. 두 방법에 의해 실험한 데이터베이스들을 평가한 결과, 동일한 영상에 대해 NFIQ와 hybrid segmentation의 결과가 유사하며 지문 영상의 품질을 세분화하여 측정할 수 있는 점에서 NFIQ보다 뛰어나다고 할 수 있다.

The purpose of this paper is to present a new measure for fingerprint image quality assessment that has a considerable effect on evaluation of fingerprint databases. This paper introduces a hybrid segmentation method for measuring an image quality and evaluates the experimental results using various fingerprint databases. This study compares the performance of the proposed hybrid segmentation using variance and coherence of fingerprints against the NIST's NFIQ program. Although NFIQ is a most widely used tool, it classifies the image quality into 5 levels. However, the proposed hybrid method is developed to be conformant to the ISO standards and accordant to human visual perception. The experimental results demonstrate that the hybrid method is able to produce finer quality measures.

In this paper, an accurate segmentation algorithm of video dynamic background image (VDBI) based on improved wavelet transform is proposed. Based on the smooth processing of VDBI, the traditional wavelet transform process is improved, and the two-layer decomposition of dynamic image is realized by using two-dimensional wavelet transform. On the basis of decomposition results and information enhancement processing, image features are detected, feature points are extracted, and quantum ant colony algorithm is adopted to complete accurate segmentation of the image. The maximum SNR of the output results of the proposed algorithm can reach 73.67 dB, the maximum time of the segmentation process is only 7 seconds, the segmentation accuracy shows a trend of decreasing first and then increasing, and the global maximum value can reach 97%, indicating that the proposed algorithm effectively achieves the design expectation.

Current methods for semantic segmentation of remote-sensing images, especially for irregular and small targets, often result in low precision and incomplete feature extraction. To address this issue, an improved semantic segmentation method was developed utilizing DeepLabv3+. First, DeepLabv3+ is combined with the proposed feature fusion module to make full use of the complementary information of low- and high-level features. Second, the channel attention module helps extract effective features while suppressing irrelevant features, thereby enabling the extraction of more meaningful global information from high-level features. Finally, rich spatial information is selected using guided spatial attention, which improves the accuracy of edge segmentation of target objects. The results of the comparison show that the mean F1 score (MF1) and overall accuracy (OA) of the proposed method on the ISPRS Potsdam dataset are 89.81% and 88.45%, respectively. The MF1 of the proposed method is 89.90% and the OA is 89.14% for the UAVid dataset, which are higher than those of the other comparison algorithms. The proposed method exhibits superior semantic segmentation capabilities for remote-sensing images.

Colonoscopy is the most effective examination way to detect colon polyps, which are highly related to colorectal cancer. Consequently, it is an important step to segment the poly accurately for diagnosis in clinical practice. However, most prior works focus on performance improvement using deep convolutional neural networks while the discrepancy between the training dataset and the test dataset is ignored. These distribution discrepancies may lead to the model overfitting the training dataset and lacking generalizability on unseen target domains. To alleviate this issue, we propose a Randomized Local Illumination Enhancement Network for polyp image segmentation. Specifically, we first employ an illumination decomposition network to decompose the input images into an illumination component and a reflectance component. The illumination component is augmented by randomly selected local illumination. Then the randomized local illumination-enhanced images are obtained by combining the augmented illumination and the reflectance, which are fed as the input of the segmentation network for improving the model generalizability. We conduct both quantitative and qualitative experiments on four polyp segmentation datasets. The satisfying results demonstrate the effectiveness of our proposed approach in the improvement of model generalizability on unseen data.