http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
함현식(Hyun-sik Ham),김동현(Dong-hyun Kim),채정우(Jung-woo Chae),이신애(Sin-ae Lee),김윤지(Yun-ji Kim),조현욱(Hyun Uk Cho),조현종(Hyun-chong Cho) 대한전기학회 2020 전기학회논문지 Vol.69 No.2
The early detection of plant disease is important in that it enhances the quality and productivity of crops. A large amount of research has considered machine learning classifiers to protect tomato plants from diseases, but the reliability of early disease diagnoses in this way remains uncertain due to the use of small datasets. Therefore, to enhance the dependability of them, this study examined a tomato disease classification system based on a deep learning using a dataset containing 17,063 images of tomato leaves infected with eight diseases. The deep learning model used in this classifier consisted of symmetric and asymmetric building blocks including convolutions, average pooling, max pooling, concats, dropouts, and fully connected layers. The obtained result indicated a high degree of accuracy (98.9%) which is high enough to be used as a proper diagnosis tool for farmers who lack professional knowledge of tomato diseases.
특징 선택을 활용한 머신 러닝 기반의 위암 컴퓨터 보조 진단 시스템
김윤지(Yun-ji Kim),이신애(Sin-ae Lee),김동현(Dong-hyun Kim),채정우(Jung-woo Chae),함현식(Hyun-sik Ham),조현진(Hyun Chin Cho),조현종(Hyun-chong Cho) 대한전기학회 2020 전기학회논문지 Vol.69 No.1
Gastric cancer is a kind of cancer that is difficult to detect at an early stage because it has almost no symptoms at the beginning. In this study, we propose a Computer-aided Diagnosis(CADx) system that detects gastric cancer from the endoscopy. The data set we used consist of 93 normal images and 93 gastric images. We extracted 6 features in 449 dimensions from the gastric endoscopy images and reduced them to 10 dimensions through feature selection algorithms. Algorithms that we use to dimension reduction are Pearson Correlation, Chi-Squared Test, Recursive Feature Elimination, and Model-based Feature Selection, which are provided by the Sci-kit Learn library. A method was also used to select the top 10 features with a higher number of times selected by these four algorithms. Normal images and gastric cancer images were classified using support vector machine(SVM). Recursive feature elimination algorithm has the highest performance among the five feature selection algorithms, with an accuracy of 0.92.
함현식(Hyun-sik Ham),조현종(Hyun-chong Cho) 대한전기학회 2021 전기학회논문지 Vol.70 No.1
The tomato is one of important crops in the world market with high commercial value. The early detection of disease is crucial for an successful crop yield. Many studies have recently been conducted to identify plant disease. In this paper, tomato disease classification using leaf images is proposed. Using the convolutional neural network(CNN), the features of disease are extracted and learned to classify. Data augmentation methods, Google’s AutoAugment algorithm and GAN(Generative Adversarial Networks), are used to increase tomato disease data. The classification model classifies nine classes of tomato disease. We compared the original model with the data augmentation models and explored that the classification that produced good performance. As a result, the SVHN policy of AutoAugment model achieved F1 Score 0.945.
다양한 이미지 증대 기법에 따른 토마토 병충해 분류 성능 개선 연구
함현식(Hyun-Sik Ham),조현종(Hyun-Chong Cho) 대한전기학회 2021 전기학회논문지 Vol.70 No.12
Crop diseases are damaging agricultural food worldwide. Early detection of crop diseases is important to prevent damage to crops. However, it is difficult to distinguish crop diseases unless not expert. Therefore, in this paper, a crop diseases classification system that can recognize diseases even before expert judgment is proposed. The characteristics of diseases were learned and classified using Convolutional Neural Network(CNN). In addition, images are augmented to increase classification performance. To augment the image, a basic augmentation policy consisting of Random Crop and Random Horizontal Flip and Google"s AutoAugment are used. Then, the augmented image was selected as an base augmentation model by setting a threshold and then trained. We compared the performance of the original, augmentation model, and image selection model. As a result, the image selection model set to the threshold value of 0.7 in AutoAugment achieved the performance of F1 Score 0.958