Earticle

Artificial neural networks have been constantly increasing in size and complexity, so their resource demands have also increased. These high computational requirements and processing time make them impractical for real-life development scenarios involving embedded systems. Resource-constrained environments such as mobile devices, IoT gadgets and edge computing platforms demand efficient models with lower computational complexity and fast real-time inference speeds. We have developed an iterative pruning technique to reduce the inference time of the model by pruning less essential neurons. Unlike traditional pruning methods that require a separate pruning step after training, our technique prunes the network gradually as it learns. This method ensures the model adapts dynamically by removing unnecessary parameters while maintaining accuracy. Our technique works by temporarily reducing the weights of a few neurons and then studying how the networks resist those neurons. Neurons with high resistance are restored to their original state, while the others with low resistance are pruned.

최근 인공지능 딥러닝 분야는 컴퓨팅 자원의 높은 연산량과 가격문제로 인해 상용화에 어려움이 존재했다. 본 논문은 더블 프루닝 기법을 적용하여 심층신경망 모델들과 다수의 데이터셋에서의 성능을 평가하고자 한다. 더 블 프루닝은 기본의 네트워크 간소화(Network-Slimming)과 파라미터 프루닝(Parameter-Pruning)을 결합한다. 이는 기존의 학습에 중요하지 않는 매개변수를 절감하여 학습 정확도를 저해하지 않고 속도를 향상시킬 수 있다는 장점이 있다. 다양한 데이터셋 학습 이후에 프루닝 비율을 증가시켜, 모델의 사이즈를 감소시켰다. NetScore 성능 분석 결과 MobileNet-V3가 가장 성능이 높게 나타났다. 프루닝 이후의 성능은 Cifar 10 데이터셋에서 깊이 우선 합성곱 신경망으로 구성된 MobileNet-V3이 가장 성능이 높았고, 전통적인 합성곱 신경망으로 이루어진 VGGNet, ResNet또한 높은 폭으로 성능이 증가함을 확인하였다.

Recently, the artificial intelligence deep learning field has been hard to commercialize due to the high computing power and the price problem of computing resources. In this paper, we apply a double pruning techniques to evaluate the performance of the in-depth neural network and various datasets. Double pruning combines basic Network-slimming and Parameter-prunning. Our proposed technique has the advantage of reducing the parameters that are not important to the existing learning and improving the speed without compromising the learning accuracy. After training various datasets, the pruning ratio was increased to reduce the size of the model.We confirmed that MobileNet-V3 showed the highest performance as a result of NetScore performance analysis. We confirmed that the performance after pruning was the highest in MobileNet-V3 consisting of depthwise seperable convolution neural networks in the Cifar 10 dataset, and VGGNet and ResNet in traditional convolutional neural networks also increased significantly.

As the connecters of greenspace networks, urban trees in street (UTS) accumulate atmospheric carbon during growth via photosynthesis processes. However, excessive management of UTS including pruning causes negative effects to return absorbed carbon to the atmosphere. This study identified the management practices of pruning and developed regression models that easily estimates annual carbon emission by the pruning amount of each tree species. The pruning practices of UTS was surveyed targeting trees planted in Chuncheon and Daegu through maintenance data, interviews with managers and actual measurements. For measuring the pruning amount, Ginkgo biloba and Platanus occidentalis with a high relative planting frequency for which pruning is mainly carried out, were selected. The annual pruning frequency for Ginkgo biloba and Platanus occidentalis was 2 times, and the amount of gasoline consumption from operating the pruning equipment was 166.7 mL per tree. All the regression models by the pruning showed fitness with r2 values of 0.75 or larger. The pruning amount and carbon emissions of Ginkgo biloba and Platanus occidentalis increased accordance with the size. The carbon emissions from Platanus occidentalis (8.5 kg/yr, at dbh 30 to 39 cm) were larger than that from Ginkgo biloba (2.8 kg/yr, at dbh 30 to 39 cm). The results of this study can serve as useful base data necessary to quantify the net carbon uptake of UTS in the future.

Previous studies surmised that under-canopy dead branches in unmanaged/abandoned coniferous plantations could reduce throughfall kinetic energy (TKE); however, the effect of under-canopy dead branch pruning on TKE has not been clarified. We established two plots under the same stand density (2500 stems ha⁻¹) in abandoned Japanese cypress plantations: (1) one with dead branch pruning (plot 1, P1) and (2) the other with no dead branch pruning (plot 2, P2). Throughfall (TF) and TKE were measured on the weekly basis during the growing season using manual-type TF rain gauges and sand-filled splash cups, respectively. Results showed that other than dead branch structures, no other stand structures significantly differed between P1 and P2 (p < 0.05). Both stand-scale TF ratio (72.8% vs. 63.5%) and unit TKE (21.4 J m⁻² mm⁻¹ vs. 13.5 J m⁻² mm⁻¹) were higher in P1 than in P2. This implies that dead branches under the canopy had considerable impacts on raindrop fall velocity after passing through upper canopies in unmanaged/abandoned Japanese cypress plantations. Our findings clarified that under-canopy dead branches can mitigate soil erosion risk in such plantations and can provide new insights into forest management for soil conservation.

This paper presents a method to optimize 3D Gaussian Splatting, improving computational efficiency for real-time applications on edge devices like Jetson AGX Xavier. By applying pruning and quantization techniques, we enhance computing speed with minimal degradation on image quality. Our method enables efficient 3D scene reconstruction on resource-limited devices, making it suitable for AR/VR and autonomous driving.

As the number of vehicle drivers is increasing day by day, the risk of traffic accidents is also increasing. Among the accidents, there are tire-related accidents causes huge damage if it occurs. These kinds of accidents could be prevented through safety checks of tire, but drivers usually overlook it because they don’t have the knowledge to know what the condition of the tire and don’t want to spend time and money to safety inspection and so on. To solve these problems, we propose tire life prediction mobile application with deep-learning method to check the condition of tires simply. Also, considering the embedded environment that has low power and capacity, we apply lightweight technique called pruning

Background: This study were performed to determine the effect of root pruning of Zizyphus jujuba var. inermis (Bunge) Rehder. Root cutting inhibit vegetative growth and promote reproductive growth as temporarily reducing growth, net assimilation, water potential of leaf and cytokinin level. Methods and Results: The root pruning was treated of the root cutting widths 50, and 80 ㎝ and the root cutting depths 10, and 20 ㎝. The amount of root pruning and the number of suckers were the highest in the root-pruning treatment at a width of 50 ㎝ and a depth of 20 ㎝. The blooming time was from June 18 to 20, and no difference was observed in the blooming time among the rootpruning treatments. The number of flowers was rather higher in the root-pruning treatment at a width of 50 ㎝ and a depth of 20 ㎝ and at a width of 80 ㎝ and a depth of 20 ㎝. The percentage of fruit setting was higher in the plants whose roots were pruned at a depth of 20 ㎝ than in the untreated plants. The fruit size, fruit weight, and sugar content showed no difference among the root-pruning treatments. Conclusions: The results showed that percentage of fruit setting increased with root pruning, while no difference was observed in the growth and fruit quality of plants.

The purpose of this experiment is to improve the ratio of female flower in Schizandra chinensis Baillon. The experiment was carried out in two purposes. One was to understand the development stages on bud formation and determining sex after bud formation by stage. The other was to investigate characteristics of flowers forming on spring when summer pruning was taken place in different cutting part and timing in S. chinensis Baillon. The outcome was that on 7th of June the flower bud was found next to leaf primordium in branch of this year and no changes were found in shape for two weeks by 26th July. On 26th of June, an ovule that forms organ was observed and then, an ovule was found after sex was determined on 6yh of July. Second experiment was carried out to find out when the sex was determined by adapting different pruning methods; cutting the branch at 6th node on June , June 15, and July 1; cutting the branch at 9th node on June 15, and July 1. The result was that the re-growth ratio on the branch 6th node was cut on June 1 and June 15. July 1 was 34-56% and, however, the number of setting flowers and mail flowers were fewer than the normal branches. More over, even though the re-growth ratio of branches were 34-25% when the cutting was taken place at 10th node on June 15 and July 1, the number of setting flowers and mail flowers were the same as the non treated branch. As a result, it is assumed that cutting controlling in this plant to increase the number of female flowers has no effect.

가지치기 작업은 보통 묘목의 식재 후 수고가 6 m 전후일 때 1차 작업을 실시하고 수고 12~13 m 전후일 때 2차 작업을 실시한다. 옹이가 없고 통직한 목재를 생산하기 위해서는 도구나 기계에 의하여 가지치기 작업이 필요하다. 가지치기 작업방법에는 자동지타기를 이용하거나 동력 고지톱을 이용하는 등 여러 가지 방법이 있지만 본 연구에서는 가지치기 작업에 일반적으로 가장 많이 활용하고 있는 인력고지톱에 의한 가지치기작업을 조사하였다. 조사수종은 강원도 지역을 중심으로 분포하고 있는 소나무,잣나무,리기다소나무를 대상으로 하였다.조사대상 수고는 10~16 m 내외로서 가지치기 후 지하고의 높이는 6.2~6.7 m 내외로서 인력고지톱에 의한 가지치기 작업능률은 4~6 m의 작업높이의 경우 1본당 소나무 3.46분, 잣나무 5.06분, 리기다소나무 4.44분이 소요되었고, 1일 작업능률은 소나무 104본, 잣나무 64본, 리기다소나무 81본으로 나타났다.

The first pruning works of planted trees on forest area carry out when tree height reached at 6 meters. And the second works carry out when it grow to 12~13 meters of tree height. Pruning works are necessary for producing straight log without knar by tool or machine. Generally, the mechanized pruning works Self-propelled pruning machine, chain pruning saw and other tools are used in mechanized pruning works. However, manual pruning saw which is usually using pruning tool was for this study. To investigate the pruning works efficiency, Pinus densiflora, Pinus koraiensis and Pinus rigida which were distributed in Kangwon-Do was surveyed. Height of surveyed the trees were 10~16 meters and its pruning works range were 6.2~6.7 meters of tree height. As results, pruning works efficiency of Pinus densiflora, Pinus koraiensis and Pinus rigida were 3.14 min/tree, 5.06 min/tree and 4.44 min/tree, respectively. Also, possible pruning works of man-day for Pinus densiflora, Pinus koraiensis and Pinus rigida was 104, 64, and 81 trees, respectively.

본 논문에서는 에이다부스트를 위한 약한 학습기로 LDA(Linear Discriminant Analysis)를 적용하여 결정트리를 구성하면서 분류 성능을 향상하도록 하는 방안에 관해 다룬다. LDA는 데이터를 최적의 투사면에 투사하여 분포를 선형적 으로 표현하고 엔트로피를 최소화함으로써 차원 축소와 부류 간 관계를 효과적으로 반영한다. 이는 에이다부스트에서 일 반적으로 사용하는 CART 결정트리와 달리 속성 축과 관계없이 자유로운 각도로 데이터를 분류할 수 있어 더 유연한 구조 의 결정트리를 생성한다. LDA를 약한 학습기로 활용하기 위해서는 학습 중단 기준을 설정하는 절단(pruning) 알고리즘이 필요하며, 랜덤보다 약간 높은 성능을 가진 결정트리를 구성하는 데 초점이 맞춰진다. 논문에서는 절단 알고리즘을 통해 노이즈 비율에 따라 노드 수와 성능의 변화를 비교하고, 에이다부스트 알고리즘에 필요한 최적의 모델 수를 실험적으로 분석하여 성능에 미치는 영향에 대해 알아본다.

This paper explores an approach to enhance classification performance by employing Linear Discriminant Analysis (LDA) as a weak learner for AdaBoost, constructing decision trees in the process. LDA projects data onto an optimal subspace, linearly representing its distribution while minimizing entropy, thereby effectively capturing inter-class relationships and enabling dimensionality reduction. Unlike the commonly used CART decision trees in AdaBoost, LDA facilitates the creation of more flexible decision trees by classifying data at arbitrary angles, independent of attribute axes. To utilize LDA as a weak learner, a pruning algorithm is required to define a stopping criterion during training. The focus is on constructing decision trees with slightly higher performance than random guessing. The study compares changes in node count and performance relative to noise levels using the pruning algorithm, and experimentally analyzes the optimal number of models required for the AdaBoost algorithm. The impact of these factors on performance is also investigated.

This study was conducted to confirm whether domestic fruit tree pruning residues can replace sawdust, which is the main ingredient of mushroom medium. The five types of fruit tree pruning residues collected were from apples, pears, peaches, grapes, and citrus. The basic components of these residues were analyzed. The pH ranged from 5.2 to 7.0, the Total carbon(T-C) ranged from 46.0% to 47.2%, the Total nitrogen(T-N) ranged from 0.5% to 0.9%, and the moisture content ranged from 12.4% to 14.2%, which was identified as an appropriate range for growing mushrooms. In order to confirm the possibility of mycelial growth of oyster mushroom "Suhan", column tests were conducted based on the conventional medium(poplar sawdust 49.5: cotton seed hull 27.3: beet pulp 12.7: cotton seed meal 10.5, v/v). As a result of incubation at 25°C for 28 days, grape branches showed the fastest growth at 143 mm compared to the control, which showed 135 mm. The yield per bottle was comparable, with grape branches(134g/bottle) and control(139g/bottle). Additionally, the quality of the fruiting bodies was comparable across all Processed lots. It means grape branches can be used as alternative sawdust materal. In the future, it is expected that using by-products as substitutes for carbon sources and main ingredients will help reduce farm production costs and protect the environment.

Deep neural networks (DNNs) have been widely used in various applications, however, the computational complexity and memory requirements of DNNs are becoming increasingly challenging, especially in resource-constrained devices such as mobile phones and embedded systems. In this paper, we propose a lightweight DNN model using channel pruning to address the computational complexity and memory requirements of DNNs in resource-constrained devices. Our approach combines channel pruning with transfer learning to maintain accuracy. Evaluation on the CIFAR-10 dataset shows improved performance with 78% test accuracy, 89% train accuracy, and 73% validation accuracy compared to the unpruned model. The pruned model is suitable for applications with limited computational resources.

엣지 디바이스 환경에서 실시간 화재 탐지를 위해서는 높은 정확도와 함께 제한된 연산 자원 내에서의 효율적인 모 델 구동이 요구된다. 그러나 기존 YOLOv12 기반 객체 탐지 모델은 높은 연산량과 복잡도로 인해 엣지 환경 적용 에 제약이 있다. 이를 해결하기 위해 본 논문에서는 K-means clustering과 FPGM을 결합한 FPKM 기반 구조적 가지치기 방법을 제안하고, 구조적 가지치기 이후 발생할 수 있는 채널 불일치 문제를 해결하기 위한 모델 재구성 전략을 함께 적용하였다. 특히 Residual, Concatenate, Split 구조가 포함된 네트워크에서도 연산 정합성을 유지 하도록 설계하였다. 실험 결과, FPKM은 30%, 50%, 70% 전 구간에서 FLOPs와 추론 시간을 감소시키면서도 성능 저하는 제한적으로 유지하였다. GPU(RTX 2080 Ti) 환경에서 50% 가지치기 시 추론 시간은 4.4 ms에서 2.6 ms로 약 41% 단축되었으며, mAP50은 79.5%에서 77.3%로 2.2% 감소하였다. CPU 환경에서도 baseline 54.3 ms, 5.8 G FLOPs에서 32.5 ms, 3.7 G FLOPs(50%)로 감소하여 약 40% 수준의 연산 효율 개선을 확인하였다. 이를 통해 제안한 방법이 GPU뿐 아니라 CPU 기반 엣지 환경에서도 효과적인 경량화 기법임 을 확인하였다.

Real-time fire detection in edge-device environments requires high detection accuracy as well as efficient model operation under limited computational resources. However, the original YOLOv12- based object detection model has limitations for edge deployment due to its high computational complexity and model size. To address this issue, we propose an FPKM-based structural pruning method that combines K-means clustering with Filter Pruning via Geometric Median (FPGM), along with a reconstruction strategy to resolve channel mismatch problems caused by structural pruning. The reconstruction framework is designed to preserve tensor consistency and computational stability even in networks containing branching structures such as Residual, Concatenate, and Split connections. Experimental results show that the proposed FPKM method reduces both FLOPs and inference time across pruning ratios of 30%, 50%, and 70%, while maintaining only limited performance degradation. In the GPU (RTX 2080 Ti) environment, at a 50% pruning ratio, inference time was reduced from 4.4 ms to 2.6 ms (approximately 41% reduction), while mAP50 decreased from 79.5% to 77.3% (a 2.2% drop). In the CPU environment, the baseline model required 54.3 ms and 5.8 G FLOPs, whereas the 50% pruned model achieved 32.5 ms and 3.7 G FLOPs, demonstrating about 40% improvement in computational efficiency. These results confirm that the proposed method is effective not only in GPU environments but also in CPU-based edge scenarios, making it suitable for real-time fire detection on edge devices.

수레바퀴 중심축(허브, 정점)이 요구하는 살(간선)을 교차없이 연결하여 모든 중심축이 연결된 망을 형성하는 수레바퀴 살 퍼즐 문제는 연구의 불모지라 할 수 있다. 이 문제에 대해서는 지수시간이 소요되는 전수 탐색법이나 분기 한정 법조차도 제시된 알고리즘이 없는 실정이다. 본 논문은 주어진 SP에 대해 m×n의 교차 대각선을 가진 격자 그래프 를 작도하고, 잉여 간선을 전정(삭제)하는 알고리즘을 제안하였다. 제안된 알고리즘은 간선 수가 허브 요구량과 일치하는 허브의 간선을 선택하고 이와 교차하는 간선을 삭제하는 단순한 방법이다. 만약 허브 요구량을 충족하는 간선을 가진 허브가 존재하지 않으면 여유 량이 최대인 허브의 간선을 우선하여 삭제(전정)하는 전략을 채택하였다. 제안된 알고리즘 을 20개의 벤치마킹 실험 데이터에 적용한 결과 모든 문제에 대해 시행착오 회수를 최소로 하는 해를 구할 수 있음을 보였다.

The problem of the spokes puzzle(SP), which connects the spokes(edges) required by the wheel axis (hub, vertex) without intersection to form a network in which all the hubs are connected, can be said to be a wasteland of research. For this problem, there is no algorithm that presents a brute-force search or branch-and-bound method that takes exponential time. This paper proposes an algorithm to plot a lattice graph with cross-diagonal lines of m×n for a given SP and to pruning(delete) the surplus edges(spokes). The proposed algorithm is a simple way to select an edge of a hub whose number of edges matches the hub requirement and delete the edge crossing it. If there is no hub with an edge that meets the hub requirement, a strategy was adopted to preferentially delete(pruning) the edge of the hub with the maximum amount of spare. As a result of applying the proposed algorithm to 20 benchmarking experimental data, it was shown that a solution that minimizes the number of trials and errors can be obtained for all problems.

섬(정점)이 요구하는 가교(간선)를 대각선을 제외한 가로와 세로 직선 가교를 교차없이 연결하여 모든 섬들이 연결된 망을 형성하는 가교 퍼즐 문제는 관련 연구가 전혀 없는 연구의 불모지라 할 수 있다. 이 문제에 대해서는 일반적 으로 알려진 지수시간이 소요되는 전수 탐색법이나 분기한정 법조차도 제시된 알고리즘이 없는 실정이다. 본 논문은 주 어진 BP에 대해 대각선이 없는 격자 그래프를 작도하고, 불필요한 간선은 삭제하고, 필수적인 가교는 보충하여 격자 그래프 초기 해를 구하였다. 다음으로 부족한 섬 쌍 매칭을 통해 해당 경로에 부족한 간선은 추가하고, 교차되는 잉여 간선(가교)은 삭제하는 방식을 채택하였다. 제안된 알고리즘을 24개 벤치마킹 실험 데이터에 적용한 결과 모든 문제에 대해 정확한 해를 구할 수 있음을 보였다.

The problem of the bridges(Hasjiwokakero, Hasi) puzzle, which connects the bridge(edge) required by the island(vertex) without crossing the horizontal and vertical straight bridges except for the diagonal to form a connected network, is a barren ground for research without any related research. For this problem, there is no algorithm that presents a generalized exponential time brute-force or branch-and-bound method. This paper obtained the initial solution of the lattice graph by drawing a grid without diagonal lines for a given BP, removing unnecessary edges, and supplementing essential bridges. Next, through insufficient island pair path matching, the method of adding insufficient edges to the route and deleting the crossed surplus edges(bridges) was adopted. Applying the proposed algorithm to 24 benchmarking experimental data showed that accurate solutions can be obtained for all problems.

지금까지의 빈발 항목 추출에서는 FP-Tree에 대한 순회와 패턴의 탐색이 필수적인 과정이기 때문에 마이닝 데이터 를 트리에 저장하는데 공간이 필요하고 탐색하는데 CPU시간이 필요하기 마련이다. 이러한 단점을 극복하기 위하여 본 논문 에서는 조건부 FP-Tree의 의존하지 않고 트랜잭션 데이터의 각 항목들의 위치 정보를 부여하여 트랜잭션 데이터를 2차원의 위치정보 Look-Up테이블로 변환하여 시간과 공간적인 접근성을 용이하게 한다. 또한 항목과 항목의 위치에 대한 매핑배열 을 병행하여 시간 복잡도를 줄이는 방법을 고려하는 알고리즘을 제안한다. 실험 결과를 통하여 제안된 방법은 FIMI 저장소 웹 사이트에서 얻은 데이터 세트를 기반으로 많은 실행 시간과 메모리 사용을 줄일 수 있음을 보였다.

Since frequent item extraction so far requires searching for patterns and traversal for the FP-Tree, it is more likely to store the mining data in a tree and thus CPU time is required for its searching. In order to overcome these drawbacks, in this paper, we provide each item with its location identification of transaction data without relying on conditional FP-Tree and convert transaction data into 2-dimensional position information look-up table, resulting in the facilitation of time and spatial accessibility. We propose an algorithm that considers the mapping scheme between the location of items and items that guarantees the linear time complexity. Experimental results show that the proposed method can reduce many execution time and memory usage based on the data set obtained from the FIMI repository website.

데이터에 내포되어 있는 주요 정보나 지식을 추출해 내는 기계학습 방법에서 주요 이슈의 하나는 지식 표현방식이다. 여러 가지 구조로 표현될 수 있는 지식을 모델이라고 부른다. 모델에는 그 내부 구조에 따라 트리구조, 네트워크 구조, 리스트 구조, 규칙 등 다양한 구조로 나눈다. 구조의 차이는 단지 표현의 차이뿐만 아니라 그것이 갖는 문제해결 능력에도 차이가 있다. 본 논문에서는 모델을 간략화 시켜 오버피팅 문제를 해결하고 분류 능력을 향상시키는 방법을 제안한다. 모델을 단순화 시키는데 사용되는 파티션 유틸리티 기준함수 제시하고 휴리스틱을 이용하여 균형 잡힌 계층 구조를 생성하는 방법을 제안한다.

One of major issues in machine learning that extracts and acquires knowledge implicit in data is to find an appropriate way of representing it. Knowledge can be represented by a number of structures such as networks, trees, lists, and rules. The differences among these exist not only in their structures but also in effectiveness of the models for their problem solving capability. In this paper, we propose partition utility as a criterion function for clustering that can lead to simplification of the model and thus avoid overfitting problem. In addition, a heuristic is proposed as a way to construct balanced hierarchical models.