京都大学 / 0048 / 新制・課程博士 / 博士(農学) / 甲第21429号 / 農博第2307号 / 学位論文||H30||N5157(農学部図書室) / 京都大学大学院農学研究科地域環境科学専攻 / (主査)教授 近藤 直, 准教授 小川 雄一, 教授 飯田 訓久 / 学位規則第4条第1項該当 / Doctor of Agricultural Science / Kyoto University / DGAM

No description available.

In the master thesis about object detection(OD) using deep convolutional neural network(DCNN), the area of OD is being tested when being applied to infrared images(IR). In this thesis the, goal is to use both long wave infrared(LWIR) images and short wave infrared(SWIR) images taken from an airplane in order to train a DCNN to detect runways, Precision Approach Path Indicator(PAPI) lights, and approaching lights. The purpose for detecting these objects in IR images is because IR light transmits better than visible light under certain weather conditions, for example, fog. This system could then help the pilot detect the runway in bad weather. The RetinaNet model architecture was used and modified in different ways to find the best performing model. The models contain parameters that are found during the training process but some parameters, called hyperparameters, need to be determined in advance. A way to automatically find good values of these hyperparameters was also tested. In hyperparameter optimization, the Bayesian optimization method proved to create a model with equally good performance as the best performance acieved by the author using manual hyperparameter tuning. The OD system was implemented using Keras with Tensorflow backend and received a high perfomance (mAP=0.9245) on the test data. The system manages to detect the wanted objects in the images but is expected to perform worse in a general situation since the training data and test data are very similar. In order to further develop this system and to improve performance under general conditions more data is needed from other airfields and under different weather conditions.

A proposta deste trabalho foi analisar diferentes metodologias de treinamento de uma rede neural convolucional profunda (CNN) para a detecção de distorção arquitetural mamária (DA) em imagens de mamografia digital. A DA é uma contração sutil do tecido mamário que pode representar o sinal mais precoce de um câncer de mama em formação. Os sistemas computacionais de auxílio ao diagnóstico (CAD) existentes ainda apresentam desempenho insatisfatório para a detecção da DA. Sistemas baseados em CNN têm atraído a atenção da comunidade científica, inclusive na área médica para a otimização dos sistemas CAD. No entanto, as CNNs necessitam de um grande volume de dados para serem treinadas adequadamente, o que é particularmente difícil na área médica. Dessa forma, foi realizada neste trabalho, uma comparação de diferentes abordagens de treinamento para uma arquitetura CNN avaliando-se o efeito de técnicas de geração de novas amostras (data augmentation) sobre o desempenho da rede. Para isso, foram utilizadas 240 mamografias digitais clínicas. Uma das redes (CNN-SW) foi treinada com recortes extraídos por varredura em janela sobre a área interna da mama (aprox. 21600 em média) e a outra rede (CNN-SW+) contou com o mesmo conjunto ampliado por data augmentation (aprox. 345000 em média). Para avaliar o método, foi utilizada validação cruzada por k-fold, gerando-se em rodízio, 10 modelos de cada rede. Os testes analisaram todas as ROIs extraídas da mama, sendo testados 14 mamogramas por fold, e obtendo-se uma diferença estatisticamente significativa entre os resultados (AUC de 0,81 para a CNN-SW e 0,83 para a CNN-SW+). Mapas de calor ilustraram as predições da rede, permitindo uma análise visual e quantitativa do comportamento de ambos os modelos. / The purpose of this work was to analyze different training methodologies of a deep convolutional neural network (CNN) to detect breast architectural distortion (AD) in digital mammography images. AD is a subtle contraction of the breast tissue that may represent the earliest sign of a breast cancer in formation. Current Computer-Aided Detection (CAD) systems still have an unsatisfactory performance on AD detection. CNN-based systems have attracted the attention of the scientific community, including in the medical field for CAD optimization. However, CNNs require a large amount of data to be properly trained, which is particularly difficult in the medical field. Thus, in this work, different training approaches for a CNN architecture are compared evaluating the effect of data augmentation techniques on the data set. For this, 240 clinical digital mammography were used. One of the networks (CNN-SW) was trained with regions of interest (ROI) extracted by a sliding window over the inner breast area (approx 21600 on average) and the other network (CNN-SW+) had the same set enlarged by data augmentation (about 345000 on average). To evaluate the method, k-fold cross-validation was used, generating 10 instances of each model. The tests looked at all the ROIs extracted from the breast (14 mammograms per fold), and results showed a statistically significant difference between both networks (AUC of 0.81 for CNN-SW and 0.83 for CNN-SW+). Heat maps illustrated the predictions of the networks, allowing a visual and quantitative analysis of the behavior of both models.

This diploma thesis focuses on a face recognition from a video, specifically how to aggregate feature vectors into a single discriminatory vector also called a template. It examines the issue of the extremely angled faces with respect to the accuracy of the verification. Also compares the relationship between templates made from vectors extracted from video frames and vectors from photos. Suggested hypothesis is tested by two deep convolutional neural networks, namely the well-known VGG-16 network model and a model called Fingera provided by company Innovatrics. Several experiments were carried out in the course of the work and the results of which confirm the success of proposed technique. As an accuracy metric was chosen the ROC curve. For work with neural networks was used framework Caffe.

The low amount solved residential burglary crimes calls for new and innovative methods in the prevention and investigation of the cases. There were 22 600 reported residential burglaries in Sweden 2017 but only four to five percent of these will ever be solved. There are many initiatives in both Sweden and abroad for decreasing the amount of occurring residential burglaries and one of the areas that are being tested is the use of prediction methods for more efficient preventive actions. This thesis is an investigation of a potential method of prediction by using neural networks to identify areas that have a higher risk of burglaries on a daily basis. The model use reported burglaries to learn patterns in both space and time. The rationale for the existence of patterns is based on near repeat theories in criminology which states that after a burglary both the burgled victim and an area around that victim has an increased risk of additional burglaries. The work has been conducted in cooperation with the Swedish Police authority. The machine learning is implemented with convolutional long short-term memory (LSTM) neural networks with max pooling in three dimensions that learn from ten years of residential burglary data (2007-2016) in a study area in Stockholm, Sweden. The model's accuracy is measured by performing predictions of burglaries during 2017 on a daily basis. It classifies cells in a 36x36 grid with 600 meter square grid cells as areas with elevated risk or not. By classifying 4% of all grid cells during the year as risk areas, 43% of all burglaries are correctly predicted. The performance of the model could potentially be improved by further configuration of the parameters of the neural network, along with a use of more data with factors that are correlated to burglaries, for instance weather. Consequently, further work in these areas could increase the accuracy. The conclusion is that neural networks or machine learning in general could be a powerful and innovative tool for the Swedish Police authority to predict and moreover prevent certain crime. This thesis serves as a first prototype of how such a system could be implemented and used.

Facial expression recognition (FER) is still a challenging concept, and machines struggle to comprehend effectively the dynamic shifts in facial expressions of human emotions. The existing systems, which have proven to be effective, consist of deeper network structures that need powerful and expensive hardware. The deeper the network is, the longer the training and the testing. Many systems use expensive GPUs to make the process faster. To remedy the above challenges while maintaining the main goal of improving the accuracy rate of the recognition, we create a generic hierarchical structure with variable settings. This generic structure has a hierarchy of three convolutional blocks, two dropout blocks and one fully connected block. From this generic structure we derived four different network structures to be investigated according to their performances. From each network structure case, we again derived six network structures in relation to the variable parameters. The variable parameters under analysis are the size of the filters of the convolutional maps and the max-pooling as well as the number of convolutional maps. In total, we have 24 network structures to investigate, and six network structures per case. After simulations, the results achieved after many repeated experiments showed in the group of case 1; case 1a emerged as the top performer of that group, and case 2a, case 3c and case 4c outperformed others in their respective groups. The comparison of the winners of the 4 groups indicates that case 2a is the optimal structure with optimal parameters; case 2a network structure outperformed other group winners. Considerations were done when choosing the best network structure, considerations were; minimum accuracy, average accuracy and maximum accuracy after 15 times of repeated training and analysis of results. All 24 proposed network structures were tested using two of the most used FER datasets, the CK+ and the JAFFE. After repeated simulations the results demonstrate that our inexpensive optimal network architecture achieved 98.11 % accuracy using the CK+ dataset. We also tested our optimal network architecture with the JAFFE dataset, the experimental results show 84.38 % by using just a standard CPU and easier procedures. We also compared the four group winners with other existing FER models performances recorded recently in two studies. These FER models used the same two datasets, the CK+ and the JAFFE. Three of our four group winners (case 1a, case 2a and case 4c) recorded only 1.22 % less than the accuracy of the top performer model when using the CK+ dataset, and two of our network structures, case 2a and case 3c came in third, beating other models when using the JAFFE dataset. / Electrical and Mining Engineering