• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.22 no.9
• /
• pp.1165-1171
• /
• 2018

CNN requires a large amount of computation and memory in the process of extracting the feature of the object. Also, It is trained from the network that the user has configured, and because the structure of the network is fixed, it can not be modified during training and it is also difficult to use it in a mobile device with low computing power. To solve these problems, we apply a pruning method to the pre-trained weight file to reduce computation and memory requirements. This method consists of three steps. First, all the weights of the pre-trained network file are retrieved for each layer. Second, take an absolute value for the weight of each layer and obtain the average. After setting the average to a threshold, remove the weight below the threshold. Finally, the network file applied the pruning method is re-trained. We experimented with LeNet-5 and AlexNet, achieved 31x on LeNet-5 and 12x on AlexNet.

• Journal of Korea Multimedia Society
• /
• v.9 no.5
• /
• pp.624-634
• /
• 2006

This paper presents a new watermarking scheme that is robust against geometrical attacks such as translation and rotation. The proposed method is based on the conventional PSADT(Polar Coordinates Shape Adaptive Discrete Transform) method which is an robust watermarking scheme for an arbitrarily-shaped image such as character images. The PSADT method shows perfect robustness against geometrical attack if there is no change in the shape of the image object. However, it cannot be utilized to watermark general rectangular images because of the missing alignment between the watermarked signals in the embedding and extracting side. To overcome this problem we propose a new watermarking scheme that aligns the watermark signal using the image inherent feature, especially corner. Namely the proposed method decides a consistent target region whose shape and position isn't changed by any malicious attack and then embeds the watermark in it using the PSADT method. Experimental results show the robustness of the proposed method against geometrical attacks as well as image compression.

• KIPS Transactions on Software and Data Engineering
• /
• v.8 no.10
• /
• pp.403-410
• /
• 2019

Image de-snowing aims at eliminating the negative influence by snow particles and improving scene understanding in images. In this paper, a perceptual generative adversarial network based a single image snow removal method is proposed. The residual U-Net is designed as a generator to generate the snow free image. In order to handle various sizes of snow particles, the inception module with different filter kernels is adopted to extract multiple resolution features of the input snow image. Except the adversarial loss, the perceptual loss and total variation loss are employed to improve the quality of the resulted image. Experimental results indicate that our method can obtain excellent performance both on synthetic and realistic snow images in terms of visual observation and commonly used visual quality indices.

• Journal of the Korea Society of Computer and Information
• /
• v.26 no.11
• /
• pp.173-182
• /
• 2021

In this paper, we analyzed driver's and passenger's motions that cause driver's distraction, and recognized 10 driver's behaviors related to mobile phones. First, distraction-inducing behaviors were classified into environments and factors, and related recent papers were analyzed. Based on the analyzed papers, 10 driver's behaviors related to cell phones, which are the main causes of distraction, were recognized. The experiment was conducted based on about 100,000 image data. Features were extracted through SURF and tested with three models (CNN, ResNet-101, and improved ResNet-101). The improved ResNet-101 model reduced training and validation errors by 8.2 times and 44.6 times compared to CNN, and the average precision and f1-score were maintained at a high level of 0.98. In addition, using CAM (class activation maps), it was reviewed whether the deep learning model used the cell phone object and location as the decisive cause when judging the driver's distraction behavior.

• Journal of IKEEE
• /
• v.23 no.3
• /
• pp.910-916
• /
• 2019

Non-local similarity of natural images is one of highly exploited features in various applications dealing with images. Unique edges, texture, and pattern of the images are frequently repeated over the entire image. Once the similar image blocks are classified into a cluster, representative features of the image blocks can be extracted from the cluster. The bigger the size of the cluster is the better the additive white noise can be separated. Denoising is one of major research topics in the image processing field suppressing the additive noise. In this paper, a denoising algorithm is proposed which first clusters the noisy image blocks based on similarity, extracts the feature of the cluster, and finally recovers the original image. Performance experiments with several images under various noise strengths show that the proposed algorithm recovers the details of the image such as edges, texture, and patterns while outperforming the previous methods in terms of PSNR in removing the additive Gaussian noise.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.24 no.3
• /
• pp.341-348
• /
• 2020

The ability to extract useful information from an image, such as the human eye, is an interface technology essential for AI computer implementation. The building recognition technology has a lower recognition rate than other image recognition technologies due to the various building shapes, the ambient noise images according to the season, and the distortion by angle and distance. The computer vision based building recognition algorithms presented so far has limitations in discernment and expandability due to manual definition of building characteristics. This paper introduces the deep learning CNN (Convolutional Neural Network) model, and proposes new method to improve the recognition rate even by changes of building images caused by season, illumination, angle and perspective. This paper introduces the partial images that characterize the building, such as windows or wall images, and executes the training with whole building images. Experimental results show that the building recognition rate is improved by about 14% compared to the general CNN model.

• Journal of the Korea Society of Computer and Information
• /
• v.27 no.10
• /
• pp.1-9
• /
• 2022

In this paper, we propose the Deep Learning-Based Companion Animal Abnormal Behavior Detection Service, which using video and sensor data. Due to the recent increase in households with companion animals, the pet tech industry with artificial intelligence is growing in the existing food and medical-oriented companion animal market. In this study, companion animal behavior was classified and abnormal behavior was detected based on a deep learning model using various data for health management of companion animals through artificial intelligence. Video data and sensor data of companion animals are collected using CCTV and the manufactured pet wearable device, and used as input data for the model. Image data was processed by combining the YOLO(You Only Look Once) model and DeepLabCut for extracting joint coordinates to detect companion animal objects for behavior classification. Also, in order to process sensor data, GAT(Graph Attention Network), which can identify the correlation and characteristics of each sensor, was used.

• KIPS Transactions on Software and Data Engineering
• /
• v.12 no.2
• /
• pp.83-90
• /
• 2023

Blurred image is an important factor in lowering image recognition rates in Computer vision. This mainly occurs when the camera is unstablely out of focus or the object in the scene moves quickly during the exposure time. Blurred images greatly degrade visual quality, weakening visibility, and this phenomenon occurs frequently despite the continuous development digital camera technology. In this paper, it replace the modified building module based on the Deep multi-patch neural network designed with convolution neural networks to capture details of input images and Attention techniques to focus on objects in blurred images in many ways and strengthen the image. It measures and assigns each weight at different scales to differentiate the blurring of change and restores from rough to fine levels of the image to adjust both global and local region sequentially. Through this method, it show excellent results that recover degraded image quality, extract efficient object detection and features, and complement color constancy.

• Journal of the Society of Disaster Information
• /
• v.19 no.4
• /
• pp.968-975
• /
• 2023

Purpose: We would like to confirm that the false positive rate of flames/smoke is high when detecting fires. Propose a method and dataset to recognize and classify fire situations to reduce the false detection rate. Method: Using the video as learning data, the characteristics of the fire situation were extracted and applied to the classification model. For evaluation, the model performance of Yolov8 and Slowfast were compared and analyzed using the fire dataset conducted by the National Information Society Agency (NIA). Result: YOLO's detection performance varies sensitively depending on the influence of the background, and it was unable to properly detect fires even when the fire scale was too large or too small. Since SlowFast learns the time axis of the video, we confirmed that detects fire excellently even in situations where the shape of an atypical object cannot be clearly inferred because the surrounding area is blurry or bright. Conclusion: It was confirmed that the fire detection rate was more appropriate when using a video-based artificial intelligence detection model rather than using image data.

• The Journal of the Acoustical Society of Korea
• /
• v.24 no.8
• /
• pp.498-504
• /
• 2005

We introduce a Korean speech recognition platform (ECHOS) developed for education and research Purposes. ECHOS lowers the entry barrier to speech recognition research and can be used as a reference engine by providing elementary speech recognition modules. It has an easy simple object-oriented architecture, implemented in the C++ language with the standard template library. The input of the ECHOS is digital speech data sampled at 8 or 16 kHz. Its output is the 1-best recognition result. N-best recognition results, and a word graph. The recognition engine is composed of MFCC/PLP feature extraction, HMM-based acoustic modeling, n-gram language modeling, finite state network (FSN)- and lexical tree-based search algorithms. It can handle various tasks from isolated word recognition to large vocabulary continuous speech recognition. We compare the performance of ECHOS and hidden Markov model toolkit (HTK) for validation. In an FSN-based task. ECHOS shows similar word accuracy while the recognition time is doubled because of object-oriented implementation. For a 8000-word continuous speech recognition task, using the lexical tree search algorithm different from the algorithm used in HTK, it increases the word error rate by $40\%$ relatively but reduces the recognition time to half.