Earticle

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.

본 논문은 지식 증류(Knowledge Distillation, KD)의 개념, 최신 동향 및 응용을 소개한다. KD는 대형 모델의 지식을 소형 모델에 전 달해 성능을 유지하면서도 모델을 경량화하는 기술이다. 수식적으로는 소프트 타겟을 기반으로 한 손실 함수 설계가 핵심이며, 로짓 (모델이 예측한 결과값) 기반, 특징 기반, 관계 기반 지식이 주요 전달 형태이다. KD는 오프라인, 온라인, 자기 증류 등 다양한 방식으로 구현되며, 컴퓨터 비전에서는 이미지 분류, 객체 검출, 영상 분할 분야에 적용된다. 특히 구조적 정보나 어텐션 맵을 활용한 기법들이 좋은 성능을 보이고 있다. 자연어 처리 분야에서는 DistilBERT, TinyBERT 등으로 대표되며, 최근에는 API 기반 대화형 LLM에 대한 블랙박스 증류도 활발하다. 한계점으로는 교사-학생 간 용량 차이, 지식 편향 전이, 데이터 의존성 등이 있으며, 이를 극복하기 위한 보 조 교사, 데이터프리 증류, 이론적 분석 등이 향후 과제로 제시된다. KD는 모델 압축과 AI 기술 확산에 있어 실용성과 학술적 중요성을 동시에 지닌 기술이다.

This paper presents an overview of the concept, recent trends, and applications of Knowledge Distillation (KD). KD is a model compression technique that transfers knowledge from a large teacher model to a smaller student model while maintaining performance. The core mechanism involves designing a loss function based on soft targets, with knowledge typically transferred in the form of logits, feature representations, or relational information. KD can be implemented in various forms such as offline, online, and self-distillation, and has been widely applied in computer vision tasks including image classification, object detection, and semantic segmentation—often leveraging structural information or attention maps for improved performance. In natural language processing, representative models include DistilBERT and TinyBERT, with recent advancements focusing on black-box distillation from API-based conversational LLMs. Challenges such as teacher-student capacity gaps, biased knowledge transfer, and data dependency remain, and future directions include using auxiliary teachers, data-free distillation, and theoretical analyses. KD holds both practical utility and academic significance in advancing model efficiency and democratizing AI deployment.

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.

In this paper, we propose a new method for generating convolution filters of a convolutional neural network (CNN) model as linear combinations of only a few basis filters that are provided as input features. In our approach, best coefficients of the linear combinations are searched (trained) with the given input basis filters (IBFs) to reconstruct the convolution filter parameters. Since all the convolution filters can be generated by the linear combinations of the IBFs, the size of a CNN model can be compressed if the number of coefficients for the linear combinations is less than that of filter parameters. Our primary goal is to investigate the possibility of expressing filters with a small set of IBFs by linear combinations. The second goal is to compress a model so that it can be beneficial when the model is distributed and stored (particularly downloaded to mobile devices through Wi-Fi).

최근 인간과 기계 간의 상호작용에 대한 연구가 활발히 이뤄지고 있으며, 언어에 담긴 정보뿐 아니라 음성에 내포된 감정을 인식하기 위해 딥러닝 모델이 적용되고 있다. 딥러닝을 활용한 음성언어 감정 인식 기술은 주로 모바일 및 IoT기기, 그리고 임베디드 시스템과 같은 제한된 컴퓨팅 자원 환경에 적용된다. 딥러닝 모델의 계층을 깊게 쌓아 정확도를 높일수록 그 연산량과 크기가 증가하기 때문에 제한된 컴퓨팅 자원으로 구동하는데 어려움이 발생한다. 본 논문에서는 가중치 가지치기와 가중치 군집화, 두 가지 기법을 각각 음성언어 감정 인식을 위한 딥러닝 모델에 적용하고 데스크탑 PC와 임베디드 시스템 환경에서 정확도와 추론 시간을 검증했다. 실험에 사용한 임베디드 보드는 NVIDIA사의 Jetson AGX Xavier이다. 가중치 가지치기 기법은 데스크탑 PC 환경에서 18.55%, 임베디드 보드 환경에서 17.84%, 가중치 군집화 기법은 데스크탑 PC 환경에서 15.32%, 임베디드 보드 환경에서 15.08%만큼 추론 시간을 개선했으며, 정확도는 두 기법 모두 기준 모델과 큰 차이가 없음을 확인하였다.

모델 인버전은 원본 학습 데이터 없이, 사전 학습된 모델로부터 반복적인 최적화를 통해 합성 입력을 복원하는 데이 터 없는 학습에서 널리 사용되는 기법이다. 그러나 최신 비전 트랜스포머에 이를 적용할 경우, 고비용의 셀프 어텐 션 메커니즘으로 인해 큰 계산적 부담이 발생하게 된다. 이를 중요하지 않은 패치들을 모두 제거함으로써 효율성을 향상시키는 희소 모델 인버전이 제안되었다. 하지만 데이터가 없는 상황에서 검증 데이터의 부재로 인한 학습 불안 정성의 증폭은 여전히 해결해야 할 문제로 남아 있다. 검증 데이터가 없는 환경에서는 모델의 정확도가 불확실해지 고, 변동성이 커지므로, 모델의 강건성 향상이 필수적이다. 본 논문에서는 데이터 없는 환경에서 생성되는 이미지의 품질과 다양성을 일관되게 유지하여, 정확도에 대한 표준편차를 낮추고 강건성을 향상시키는 방법을 제안한다. 제안 한 Adaptive AEM은 패치 제거 이후의 중요도를 재조정해 엔트로피 최소화를 촉진시킨다. 실험 결과, 제안한 방법 으로 생성된 이미지를 사용하면 이전 방법론에 비해 데이터 없는 양자화에서는 최대 72%, 데이터 없는 지식 증류 에서는 최대 49%까지 정확도의 표준편차를 줄여 모델을 강건하게 만들 수 있음을 입증한다.

Model inversion is a widely used technique in data-free learning, where synthetic inputs are reconstructed from a pretrained model through iterative optimization without access to the original training data. However, when applied to modern Vision Transformers, the high computational cost of the self-attention mechanism poses a significant challenge. Sparse Model Inversion (SMI) has been proposed to improve efficiency by removing non-essential patches. Nevertheless, in the absence of real validation data, the instability of training remains an unresolved issue, as model accuracy becomes uncertain and exhibits high variance. To address this, we propose a method that consistently preserves the quality and diversity of generated images in data-free environments, thereby reducing the standard deviation of accuracy and enhancing model robustness. The proposed Adaptive AEM readjusts the importance after patch removal to promote entropy minimization. Experimental results demonstrate that using images generated by our method reduces the standard deviation of accuracy by up to 72% in data-free quantization and up to 49% in data-free knowledge distillation, compared to previous approaches, leading to significantly more robust models.

Purpose: The purpose of the study is to evaluate the distance and location of the rescuer to patient for the effective chest compressions qualities. Methods: The subjects were 42 students who earned the basic lifesaving technique and had informed consents to participate in the study from May 1 to 20 in 2013. The position of the rescuers included model-0(reference point), model-1(10 cm distance), model-2(20 cm distance), and model-kn(kneeling up). Results: The mean depth of compression was 50.6±6.6 mm in Model-0, 48.7±8.2 mm in Model-1, 44.2±10.4 mm in Model-2, and 51.8±6.0 mm in Model-kn. There were statistically significant differences between each Model(p<.001). Conclusion: The closer distance between rescuer and patient could provide more effective chest compressions. Kneeling on the bed stance provided the deeper chest compression consistently than the stool stance.

On-device AI environments require real-time processing but are constrained by limited computational resources. Previous studies have shown that simply replacing high-cost computational modules with low-cost alternatives does not always yield actual speed improvements on embedded hardware. Therefore, this study aims to design a YOLOv8-n–based lightweight network that can achieve real-time inference and high accuracy under stringent resource constraints. The proposed model, LGO-YOLO, applies module structures optimized for embedded computation to both the Backbone and Neck, reducing the model’s computational load and number of parameters by approximately 42% and 40%, respectively. Despite these reductions, the model achieves accuracy and precision equal to or superior to YOLOv8-n in several performance metrics—most notably, an mAP@0.5 of 99.3%. Furthermore, in an NPU environment, it records the fastest inference time (25.4 ms) among all comparison models. This work demonstrates how careful structural design can balance the limits of model lightweighting with performance requirements, indicating that the proposed network can be effectively deployed in real embedded systems or other low-power application scenarios.

In this paper, a new approach is proposed to investigate the characteristics behavior of concrete under uniaxial and biaxial compression using the theory of plasticity. This approach is based on artificial neural networks (ANNs), especially radial basis function (RBF) in conjunction with the models of theory of plasticity. The main advantage of the proposed approach is to estimate the quality of the results with accuracy equivalent to the experiments. Another advantage of the proposed ANNs models are that it takes into account the uniaxial as well as the biaxial compression strain. The proposed models were evaluated against several experimental results available in the open literature for the behavior of the force and deformation of the two types of compression tests. Good agreement has been found between our models and those presented elsewhere.

Image compression-decompression methods have become increasingly crucial in modern times, facilitating the transfer of high-quality images while minimizing file size and internet traffic. Historically, early image compression relied on rudimentary codecs, aiming to compress and decompress data with minimal loss of image quality. Recently, a novel compression framework leveraging colorization techniques has emerged. These methods, originally developed for infusing grayscale images with color, have found application in image compression, leading to colorization-based coding. Within this framework, the encoder plays a crucial role in automatically extracting representative pixels—referred to as color seeds—and transmitting them to the decoder. The decoder, utilizing colorization methods, reconstructs color information for the remaining pixels based on the transmitted data. In this paper, we propose a novel approach to image compression, wherein we decompose the compression task into grayscale image compression and colorization tasks. Unlike conventional colorization-based coding, our method focuses on the colorization process rather than the extraction of color seeds. Moreover, we employ the Denoising Diffusion Null-Space Model (DDNM) for colorization, ensuring high-quality color restoration and contributing to superior compression rates. Experimental results demonstrate that our method achieves higher-quality decompressed images compared to standard JPEG and JPEG2000 compression schemes, particularly in high compression rate scenarios.

A new distributed algorithm of data ccompression based on hierarchical cluster model for sensor networks s proposed, the basic ideas of which are as follows, firstly the whole sensor network is mapped into a kind f hierarchical clusters model, and then different wavelet transform models are used to commit data ompression in inner and super clusters respectively, according to the relative regularity of sensor nodes eployed in the inner clusters, and the relative irregularity of sensor nodes deployed in super cluster. heoretical analyses and simulation results show that, the above new methods have good performance of pproximation, and can compress data and reduce the amount of data efficiently. So, it can prolong the ifetime of the whole sensor network to a greater degree.

딥 러닝의 발전으로 다양한 AI 기반의 응용이 많아지고, 그 모델의 규모도 매우 커지고 있다. 그러나 임베디드 기기와 같이 자원이 제한적인 환경에서는 모델의 적용이 어렵거나 전력 부족 등의 문제가 존재한다. 이를 해결하기 위해 서 클라우드 기술 또는 오프로딩 기술을 활용하거나, 모델의 매개변수 개수를 줄이거나 계산을 최적화하는 등의 경량화 방법이 제안되었다. 본 논문에서는 다양한 프레임워크들의 상호 교환 포맷으로 사용되고 있는 ONNX(개방형 신경망 교환 포맷) 포맷에 딥러닝 경량화 방법 중 학습된 모델의 양자화를 적용한다. 경량화 전 모델과의 신경망 구조와 추론 성능을 비교하고, 양자화를 위한 다양한 모듈 방식를 분석한다. 실험을 통해 ONNX의 양자화 결과, 정확도는 차이가 거의 없으며 기존 모델보다 매개변수 크기가 압축되었으며 추론 시간 또한 전보다 최적화되었음을 알 수 있었다.

Due to the development of deep learning and AI, the scale of the model has grown, and it has been integrated into other fields to blend into our lives. However, in environments with limited resources such as embedded devices, it is exist difficult to apply the model and problems such as power shortages. To solve this, lightweight methods such as clouding or offloading technologies, reducing the number of parameters in the model, or optimising calculations are proposed. In this paper, quantization of learned models is applied to ONNX models used in various framework interchange formats, neural network structure and inference performance are compared with existing models, and various module methods for quantization are analyzed. Experiments show that the size of weight parameter is compressed and the inference time is more optimized than before compared to the original model.

The compression of CAD models is a key technology for realizing Internet-based collaborative product development because big model sizes often prohibit us to achieve a rapid product information transmission. Although there exist some algorithms for compressing discrete CAD models, original precise CAD models are focused on in this paper. Here, the characteristics of hierarchical structures in CAD models and the distribution of their redundant data are exploited for developing a novel data encoding method. In the method, different encoding rules are applied to different types of data. Geometric data is a major concern for reducing model sizes. For geometric data, the control points of B-spline curves and surfaces are compressed with the second-order predictions in a local coordinate system. Based on analysis to the distortion induced by quantization, an efficient method for computation of the distortion is provided. The results indicate that the data size of CAD models can be decreased efficiently after compressed with the proposed method.

The present study investigated the role of ERK in the onset of mechanical and cold allodynia in a rat model of compression of the trigeminal ganglion by examining changes in the air-puff thresholds and number of scratches following the intracisternal injection of PD98059, a MEK inhibitor. Male Sprague Dawley rats weighing between 250 and 260 g were used. Under anesthesia, the rats were mounted onto a stereotaxic frame and received 4% agar ($10\;{\mu}l$) solution to compress the trigeminal ganglion. In the control group, the animals were given a sham operation without the application of agar. Changes in behavior were examined at 3 days before and at 3, 7, 10, 14, 17, 21, 24, 30, and 40 days after surgery. Compression of the trigeminal ganglion significantly decreased the air-puff thresholds. Mechanical allodynia was established within 3 days and persisted over postoperative day 24. To evaluate cold allodynia, nociceptive scratching behavior was monitored after acetone application on the vibrissa pad of the rats. Compression of the trigeminal ganglion was found to produce significant cold allodynia, which persisted for more than 40 days after surgery. On postoperative day 14, the intracisternal administration of $1\;{\mu}g$ or $10\;{\mu}g$ of PD98059 in the rat model significantly decreased the air-puff thresholds on both the ipsilateral and contralateral side. The intracisternal administration of $10\;{\mu}g$ of PD98059 also significantly alleviated the cold allodynia, compared with the vehicle-treated group. These results suggest that central ERK plays an important role in the development of mechanical and cold allodynia in rats with compression of the trigeminal ganglion and that a targeted blockade of this pathway is a potential future treatment strategy for trigeminal neuralgia-like nociception.

The present study investigated the role of ERK in the onset of mechanical and cold allodynia in a rat model of compression of the trigeminal ganglion by examining changes in the air-puff thresholds and number of scratches following the intracisternal injection of PD98059, a MEK inhibitor. Male Sprague Dawley rats weighing between 250 and 260 g were used. Under anesthesia, the rats were mounted onto a stereotaxic frame and received 4% agar (10 μl) solution to compress the trigeminal ganglion. In the control group, the animals were given a sham operation without the application of agar. Changes in behavior were examined at 3 days before and at 3, 7, 10, 14, 17, 21, 24, 30, and 40 days after surgery. Compression of the trigeminal ganglion significantly decreased the air-puff thresholds. Mechanical allodynia was established within 3 days and persisted over postoperative day 24. To evaluate cold allodynia, nociceptive scratching behavior was monitored after acetone application on the vibrissa pad of the rats. Compression of the trigeminal ganglion was found to produce significant cold allodynia,which persisted for more than 40 days after surgery. On postoperative day 14, the intracisternal administration of 1 μg or 10 μg of PD98059 in the rat model significantly decreased the air-puff thresholds on both the ipsilateral and contralateral side. The intracisternal administration of 10 μg of PD98059 also significantly alleviated the cold allodynia,compared with the vehicle-treated group. These results suggest that central ERK plays an important role in the development of mechanical and cold allodynia in rats with compression of the trigeminal ganglion and that a targeted blockade of this pathway is a potential future treatment strategy for trigeminal neuralgia-like nociception.

The beneficial effects of graduated compression stockings (GCS) in prophylaxis and treatment of venous disorders of human lower extremity have been recognized. However, their pressure functional performances are variable and unstable in practical applications, and the exact mechanisms of action remain controversial. Direct surface pressure measurements and indirect material properties testing are not enough for fully understanding the interaction between stocking and leg. A three dimensional (3D) biomechanical mathematical model for numerically simulating the interaction between leg and GCS in dynamic wear was developed based on the actual geometry of the female leg obtained from 3D reconstruction of MR images and the real size and mechanical properties of the compression stocking prototype. The biomechanical solid leg model consists of bones and soft tissues, and an orthotropic shell model is built for the stocking hose. The dynamic putting-on process is simulated by defining the contact of finite relative sliding between the two objects. The surface pressure magnitude and distribution along the different height levels of the leg and stress profiles of stockings were simulated. As well, their dynamic alterations with time processing were quantitatively analyzed. Through validation, the simulated results showed a reasonable agreement with the experimental measurements, and the simulated pressure gradient distribution from the ankle to the thigh (100:67:30) accorded with the advised criterion by the European committee for standardization. The developed model can be used to predict and visualize the dynamic pressure and stress performances exerted by compression stocking in wear, and to optimize the material mechanical properties in stocking design, thus, helping us understand mechanisms of compression action and improving medical functions of GCS.

본 연구에서는 CNN 모델의 경량화 및 on-device 추론 가속을 목표로 한다. 경량화 기법으로는 QAT 기법을 사용하며 여러 환경에서의 성능을 비교한다. 이어서 on-device 추론 가속을 위해 Jetson Nano Board 에서 TensorRT 변환을 통해 모델을 최적화한다.

온-디바이스 머신러닝은 비용 효율성, 데이터 프라이버시, 응답성 측면에서의 강점으로 인해 점차 인기를 얻고 있다. 그러나 소형 임베디드 시스템에서는 메모리 용량이 제한적이기 때문에 심층 신경망 모델을 처리하기 쉽지 않다. 이전 연구에서는 양자화나 가지치기와 같은 다양한 모델 압축 기법들을 제안하였다. 그러나 이러한 기법들은 일반적으로 압축으로 인한 정확도 손실을 최소화하기 위해 적절한 데이터 샘플을 사용한 세심한 미세 조정을 필요로 한다. 본 연구에서는 유사한 컨볼루션 커널을 클러스터링하고 가지치기하여 입력 모델을 압축하는 새로운 훈련 후 모델 압축 방법을 제안한다. 본 연구에서 제안된 방법은 커널 간의 유사성만을 고려하므로 데이터 샘플을 필요로하지 않는다. 본 연구는 대표적인 신경망 모델을 사용하여 제안된 방법을 평가하고 적은 정확도 손실로도 메모리 사용량을 효과적으로 줄일 수 있다는 것을 입증한다.

On-device machine learning is becoming more popular for its strengths in cost efficiency, data privacy, and responsiveness. However, processing deep neural network models on small embedded systems is challenging due to their limited memory capacity. Previous work has proposed various model compression techniques, such as quantization and pruning. However, the techniques generally require careful fine-tuning with proper data samples to minimize accuracy loss from compression. This work proposes a new post-training model compression method that compresses the input model by clustering and pruning similar convolution kernels. The proposed method does not require data samples because it considers the similarity between kernels only. This work evaluates the proposed method with representative neural network models and demonstrates that the method can effectively reduce memory usage on average with small accuracy loss.

This paper introduces model compression algorithms which make a deep neural network smaller and faster for embedded systems. The model compression algorithms can be largely categorized into pruning, quantization and knowledge distillation. In this study, gradual pruning, quantization aware training, and knowledge distillation which learns the activation boundary in the hidden layer of the teacher neural network are integrated. As a large deep neural network is compressed and accelerated by these algorithms, embedded computing boards can run the deep neural network much faster with less memory usage while preserving the reasonable accuracy. To evaluate the performance of the compressed neural networks, we evaluate the size, latency and accuracy of the deep neural network, DenseNet201, for image classification with CIFAR-10 dataset on the NVIDIA Jetson Xavier.