Earticle

컴퓨터 비전(computer vision)분야의 객체인식 기술은 최근 객체인식 자체의 성능 향상과 여러 모델의 결합 등 기술이 고도화 되고 있다. 그러나 이미지 내 객체의 크기가 작은 경우에는 그 객체에 대해 얻을 수 있는 정보의 절대적인 양이 적고, 적은 범위 내부 정보가 특징(feature)을 추출하는 과정에서 왜곡이 발생하여 오히려 특징이 흐려지는 등 객체에서 얻을 수 있는 정보가 적거나 왜곡되어 저하되는 검출 성능을 개선하기 위해 기존 객체 인식 모델에 attention 기법 을 적용하여 작은 객체에 대한 정보를 더 집중적으로 추출하여 검출 성능을 높이고자한다. 객체의 검출 성능을 분석하기 위해 객체의 검출 성능을 기존 객체 인식 모델과 attention 기법 을 적용한 객체 인식 모델의 성능을 비교·분석한다. 특히 객체의 크기를 크게 세 가지(small, medium, large)로 분류 하고 모두 학습하여 모델 별로 각 크기군의 객체의 검출 성능을 비교한다.

본 연구는 실시간 객체 탐지 기술인 YOLO에 대해 YOLOv8부터 YOLO12까지의 버전과 모델에 대해 소규모 커스텀 데 이터셋 환경에서 성능평가를 수행하여, 제한된 데이터와 자원을 보유한 실제 응용 환경에서의 모델 선택 가이드라인을 제시한 다. 교통 환경 기반 6개 클래스, 2,664장, 총 6,354개의 라벨링 커스텀 데이터셋을 구축하고, Google Colaboratory(NVIDIA Tesla T4 GPU) 환경에서 동일한 학습 조건으로 YOLOv8부터 YOLO12까지 총 26개 모델을 학습·평가하였다. 연구 결과, YOLOv9m이 mAP50-95 0.520으로 자체평가 기준 중 최고 성능을 기록하였으며, 대부분의 버전에서 소형(s) 또는 중형(m) 모 델이 시간 대비 뛰어난 성능을 보였다. 대형(l) 및 초대형(x) 모델은 학습 시간이 2~3배 증가했음에도 성능은 정체되거나 감소하 였다. 공식 벤치마크와 달리 소규모 데이터셋에서는 중간 크기 이상 모델에서 성능이 감소하는 비선형적 패턴이 나타났으며, YOLO 버전보다 모델 크기 선택이 성능에 더 큰 영향을 미치는 것으로 확인되었다. 본 연구는 제한된 데이터 환경에서의 실용적 버전 및 모델 선택 기준을 제시하며, 향후 하이퍼파라미터 최적화 및 도메인이나 데이터 규모에 따른 비교 연구가 필요하다.

This study evaluates real-time object detection models in the YOLO series, from YOLOv8 to YOLO12, using a small-scale custom dataset and proposes practical guidelines for model selection in resource-constrained environments. A custom traffic-based dataset consisting of 2,664 images, 6 classes, and a total of 6,354 labeled objects was constructed. All experiments were conducted under identical training conditions using Google Colaboratory with an NVIDIA Tesla T4 GPU, and a total of 26 models from YOLOv8 to YOLO12 were trained and evaluated. The results show that YOLOv9m achieved the highest performance among the evaluated models, with an mAP50–95 of 0.520. In most versions, small (s) and medium (m) models demonstrated superior performance relative to training time. Large (l) and extra-large (x) models showed required 2–3 times longer training, while their performance remained stagnant or decreased. Unlike official benchmarks, a nonlinear performance pattern was observed in the small-scale dataset, where performance decreased in models of medium size or larger. Furthermore, model size had a greater impact on performance than the YOLO version itself. This study provides practical guidelines for selecting appropriate YOLO versions and model sizes in resource-constrained environments. Future work should explore hyperparameter optimization and extend comparative analyses across different domains and dataset scales.

Bubble defects reduce the reliability of PCB coatings. Accurate, real-time detection of small bubbles is crucial. In this study, we present a fast and lightweight YOLOv8n-based detector. It improves the average precision of small bubbles by approximately 12 points. It also increases mAP50 from 0.637 to 0.697 in 49.1 milliseconds per image. This is a practical solution that supports quality control of micro-defects. NAMAttention reweights channel and spatial features. C2F fuses backbone and neck features. The P2 layer expands the receptive field for micro-bubbles. During training, size-aware loss emphasizes small bubbles. Defect-balanced sampling addresses class imbalance. Glare synthesis improves robustness to illumination variations. During inference, Soft-NMS reduces false positives caused by overlapping boxes. Our contribution lies in the organic integration of a learning strategy specialized for micro-defects and architectural improvements. At the same time, we maintain a lightweight structure and improve detection performance and latency. We quantitatively verified the utility and practical feasibility of each component through module-bymodule erasure experiments. We also performed throughput evaluations in field-deployment scenarios. This approach is effective for PCB quality control and suitable for detecting micro-defects in similar manufacturing processes.

본 논문은 객체 위치식별 알고리즘의 성능을 향상하기 위한 레이블 재할당 방법을 제안한다. 제안한 방법은 추론 단계와 재할당 단계로 구분한다. 추론 단계에서는 학습된 모델로부터 사전 지정된 크기에 따라 다중 스케일 추론을 수행한 뒤, 이를 마스킹한 영상을 다시 한번 추론하여 강인한 클래스 종류의 추론 결과를 얻는다. 재할당 단계에서는 박스간의 IoU를 계산하여 중복 박스를 제거하고, 박스와 클래스의 빈도를 계산하여 지배적 클래스를 다시 할당하였다. 제안한 방법을 검증하기 위하여 공사현장 안전장비 인식 영상 데이터 세트에 레이블 재할당 방법을 적용하고 이를 YOLOX-L 객체 탐지 모델에서 학습하였다. 실험 결과 적용 전 대비 mAP가 3.9% 향상하여 51.07%를 달성하였으며 AP_S를 3배 이상 향상하여 14.53%를 달성하였다. 실험 결과를 통해 레이블 재할당 알고리즘이 더 우수한 성능의 모델을 훈련해 냄을 확인하였다.

In this paper, we propose a Label Reassignment Strategy to improve the performance of an object detection algorithm. Our approach involves two stages: an inference stage and an assignment stage. In the inference stage, we perform multi-scale inference with predefined scale sizes on a trained model and re-infer masked images to obtain robust classification results. In the assignment stage, we calculate the IoU between bounding boxes to remove duplicates. We also check box and class occurrence between the detection result and annotation label to re-assign the dominant class type. We trained the YOLOX-L model with the re-annotated dataset to validate our strategy. The model achieved a 3.9% improvement in mAP and 3x better performance on AP_S compared to the model trained with the original dataset. Our results demonstrate that the proposed Label Reassignment Strategy can effectively improve the performance of an object detection model.

Chimneys play a crucial role in monitoring industrial emissions and assessing air pollution. While large chimneys are easily identifiable in high-resolution satellite imagery, small chimneys pose unique challenges due to their size, arbitrary orientations, and tendency to blend into dense urban environments. Further complications arise from inconsistent annotation practices in open-source datasets, where condensing towers are labeled as chimneys, and from intra-class imbalance resulting from the underrepresentation of clustered chimneys. These clustered chimneys typically appear in parallel arrangements, and their close spatial proximity leads to non-occluded overlapping bounding boxes, hindering the model's ability to distinguish object boundaries. To address these issues, this study proposes a large chimney filtering technique based on the statistical scale distribution of chimney sizes to effectively remove condensing tower annotations. Additionally, several targeted data augmentation strategies are introduced to enhance model generalization. A comparative evaluation of six oriented object detection models was conducted. For robustness assessment, an inference dataset containing small chimneys was constructed and cross-validated using Google Earth imagery to ensure annotation accuracy. Experimental results on the artificial intelligence (AI)-Hub, BUAA-FFPP60, and DIOR-R open-source datasets demonstrate that the proposed method significantly improves small chimney detection in complex urban environments, achieving an average precision of 0.907 on the open-source datasets and 0.523 on the inference dataset. These results show good potential for improving air pollution monitoring.

Early-stage fires appear small in captured images, making accurate detection of small objects crucial for rapid response. Although convolution, pooling operations in convolutional neural networks enhance high-level target recognition, they can also lead to the loss of fine-grained details, thereby degrading the representation and detection performance of small objects. In this paper, we propose a multi-stage detection framework "U-FRCNN (Upscaled-Faster R-CNN)" that combines Faster R-CNN with image upscaling to enhance low-confidence score regions. Initially, objects are detected with Faster R-CNN; then, regions with low detection scores are upscaled and re-evaluated using the same model. Experiments on a dataset from Roboflow and manually labeled images show that our approach improves small object mAP from 0.0947 to 0.1510 (a 59.4% increase) and overall mAP from 0.2183 to 0.2483 (a 13.7% increase), while large object detection remains robust. These results demonstrate the potential of our method for effective early fire detection.

Object detection, a key area in computer vision, involves locating and classifying objects within images. The YOLO model, which combines candidate region selection, coordinate prediction, and class inference into a single stage, has set a standard for this field. Since its introduction, numerous models have been developed, with applications spanning both academia and industry. Advances in computer vision AI, sensor, and computing technologies have been foundational for autonomous driving and driver assistance systems. Specifically, traffic sign recognition systems using object detection algorithms enhance driving safety. Annotating data for these models is resource-intensive, so many rely on benchmark datasets like MS COCO. However, this dataset has limited images featuring small objects, leading to a gap in research on models capable of detecting significantly smaller objects. Our study addresses this by developing an object detection model for distant road signs as small objects, using a specialized dataset with 2,800 x 2,100 resolution images containing objects smaller than 200 x 200 pixels. We empirically evaluated various models, including State-of-the-art YOLO version and Transformer-based approaches, under different training conditions. The best-performing model for small object detection, balancing accuracy and inference speed, was the pretrained YOLO v6m, utilizing transfer learning with COCO2017 weights without self-distillation. This research contributes to small object detection technology, offering insights for developing new AI algorithms and benefiting both academic and industrial sectors.

최근 딥러닝 기반의 객체 검출 및 인식 연구가 발전해가면서 산업 및 실생활에 적용되는 범위가 넓어지고 있다. 건설 분야에도 딥러닝 기반의 시스템이 도입되고 있지만 아직은 미온적이다. 건설 도면에서 자재 산출이 수작업으로 이뤄지고 있어 많은 소요시간과 부정확한 적산 결과로 잘못된 물량산출의 거래가 생길 수 있다. 이를 해결하기 위해서 빠르고 정확한 자동 도면 인식시스템이 필요하다. 따라서 본 논문은 건설도면 내 철강 자재를 검출하고 인식하는 인공지능기반 자동 도면 인식 적산 시스템을 제안한다. 빠른 속도의 YOLOv4 기반에 소형 객체 검출성능을 향상하기 위한 복제 방식의 데이터 증강기법과 공간집중 모듈을 적용하였다. 검출한 철강 자재 영역을 문자 인식한 결과를 토대로 철강 자재를 적산한다. 실험 결과 제안한 방식은 기존 YOLOv4 대비 정확도와 정밀도를 각각 1.8%, 16% 증가시켰다. 제안된 방식의 Precision은 0.938, Recall은 1, AP_0.5는 99.4%, AP_0.5:0.95 68.8%의 향상된 결과를 얻었다. 문자 인식은 기존 데이터를 사용한 인식률 75.6%에 비해 건설도면에 사용되는 폰트에 맞는 데이터 세트를 구성하여 학습한 결과 99.9%의 인식률을 얻었다. 한 이미지 당 평균 소요시간은 검출 단계는 0.013초, 문자 인식은 0.65초, 적산 단계는 0.16초로 총 0.84초의 결과를 얻었다.

As deep learning-based object detection and recognition research have been developed recently, the scope of application to industry and real life is expanding. But deep learning-based systems in the construction system are still much less studied. Calculating materials in the construction system is still manual, so it is a reality that transactions of wrong volumn calculation are generated due to a lot of time required and difficulty in accurate accumulation. A fast and accurate automatic drawing recognition system is required to solve this problem. Therefore, we propose an AI-based automatic drawing recognition accumulation system that detects and recognizes steel materials in construction drawings. To accurately detect steel materials in construction drawings, we propose data augmentation techniques and spatial attention modules for improving small object detection performance based on YOLOv4. The detected steel material area is recognized by text, and the number of steel materials is integrated based on the predicted characters. Experimental results show that the proposed method increases the accuracy and precision by 1.8% and 16%, respectively, compared with the conventional YOLOv4. As for the proposed method, Precision performance was 0.938. The recall was 1. Average Precision AP_0.5 was 99.4% and AP_0.5:0.95 was 67%. Accuracy for character recognition obtained 99.9.% by configuring and learning a suitable dataset that contains fonts used in construction drawings compared to the 75.6% using the existing dataset. The average time required per image was 0.013 seconds in the detection, 0.65 seconds in character recognition, and 0.16 seconds in the accumulation, resulting in 0.84 seconds.

Accurate detection of small objects is critical in military surveillance and drone monitoring, yet standard detectors often fail due to feature loss during down-sampling. This study proposes a lightweight YOLOv5-based model optimized for small-object detection. The design introduces a high-resolution P2 detection head and removes the P5 head, while anchor boxes are re-optimized using a genetic algorithm aligned with dataset statistics. On a drone small-object dataset (24-32 px), the proposed model achieved +0.062 higher mAP@0.5 than YOLOv5s, +0.041 higher than YOLOv8s, and +0.020 higher than YOLOv11s, while maintaining fewer parameters (5.3M vs. 7.0-11.1 M). Additional experiments on medium-sized objects (64-96 px) confirmed strong generalization, with mAP@0.5 reaching 0.998, surpassing YOLOv5s (0.994) and YOLOv8s (0.988), and slightly exceeding YOLOv11s (0.989). Although mAP@0.5:0.95 was marginally lower compared to YOLOv8s and YOLOv11s, the proposed model achieves an effective trade-off between accuracy and efficiency, demonstrating suitability for edge deployment in defense and real-time monitoring applications.

최근 딥러닝을 기반으로 하는 컴퓨터 비전 기술의 발전으로 물체 검출 분야의 성능이 크게 향상되었다. 그러나 일반적으로 대부분의 검출 기법들이 영상 내에서 작게 표시되는 물체에 대해서는 좋은 성능을 보이지 못한다. 본 논문은 이러한 문제를 해결하기 위해 영상 내의 소형 물체를 검출하는 기법을 제안한다. 제안 기법은 딥러닝 기반의 영상분할 기법을 사용하여 소형 물체의 대략적인 위치를 추정하는 초기 영역 검출 단계와, 추정된 위치를 기반으로 물체를 검출하는 최종 검출 단계로 구성된다. 실험 결과, 본 기법은 $40{\times}40$ 이하의 소형 물체에 대하여 기존 기법 대비 재현율 측면에서 20% 이상의 성능향상을 보인다.

Recently, the performance of the object detection has been greatly improved. However, in general, most of these detection methods do not work well on small objects in the image. To solve this problem, this paper proposes a method which effectively detects small objects in the image. The proposed method consists of an initial object extraction step to estimate the approximate locations of small objects, and the refinement step to detect the target small objects within the estimated location. Experimental results support that when detecting small objects of less than $40{\times}40$, the proposed method shows higher performance of 20% above than the previous work in terms of recall rate.

EGML (Effective Gaussian Mixture Learning) 기반의 배경차분 기법을 이용한 이동객체 검출 (Moving Object Detection; MOD) 프로세서의 효율적인 하드웨어 구현 방식을 제안한다. 하드웨어 복잡도를 감소시키기 위해 배경 생성에 사용되는 일부 연산을 근사화하여 구현하였으며, 배경차분과 가우시안 계산의 나눗셈 연산에 사용되는 하드웨어 자원이 공유되도록 설계하였다. 설계한 MOD 프로세서는 MATLAB/Simulink를 이용한 HDL-netlist 시뮬레이션과 FPGA-in-the-loop 방식을 통해 기능을 검증하였다. IEEE CDW-2014 데이터 세트의 6가지 영상을 입력으로 사용하여 MOD 성능을 평가한 결과, 평균 재현율(recall)은 0.7700, 평균 정밀도(precision)는 0.7170, F-measure가 0.7293으로 평가되었다. Xilinx ISE를 이용하여 FPGA 합성한 결과, Virtex5 XC5VSX95T 디바이스에서 총 882 슬라이스와 $146{\times}36kbit$의 블록 램으로 구현되었으며, 동일한 알고리듬을 적용한 기존의 구현 사례에 비해 약 60%의 하드웨어를 감소시켰다. MOD 프로세서는 최대 75 MHz의 클록 주파수로 동작하여 $800{\times}600$ 해상도의 영상에 대해 39 fps의 성능으로 실시간 처리가 가능한 것으로 평가되었다.

This paper proposes an efficient approach for hardware implementation of moving object detection (MOD) processor using effective Gaussian mixture learning (EGML)-based background subtraction method. Arithmetic units used in background generation were implemented using LUT-based approximation to reduce hardware complexity. Hardware resources used for both background subtraction and Gaussian probability density calculation were shared. The MOD processor was verified by FPGA-in-the-loop simulation using MATLAB/Simulink. The MOD performance was evaluated by using six types of video defined in IEEE CDW-2014 dataset, which resulted the average of recall value of 0.7700, the average of precision value of 0.7170, and the average of F-measure value of 0.7293. The MOD processor was implemented with 882 slices and block RAM of $146{\times}36kbits$ on Virtex5 FPGA, resulting in 60% hardware reduction compared to conventional design based on EGML. It was estimated that the MOD processor could operate with 75 MHz clock, resulting in real-time processing of $800{\times}600$ video with a frame rate of 39 fps.