Greece

When convolutional neural networks are applied to imagesegmentation results depend greatly on the data sets used to train thenetworks. Cloud providers support multi GPU and TPU virtual machinesmaking the idea of cloud-based segmentation as service attractive. In thispaper we study the problem of building a segmentation service, whereimages would come from different acquisition instruments, by training ageneralized U-Net with images from a single or several datasets. We alsostudy the possibility of training with a single instrument and performquick retrains when more data is available. As our example we performsegmentation of Optic Disc in fundus images which is useful for glau-coma diagnosis. We use two publicly available data sets (RIM-One V3,DRISHTI) for individual, mixed or incremental training. We show thatmultidataset or incremental training can produce results that are simi-lar to those published by researchers who use the same dataset for bothtraining and validation

EANN 2020

Multidataset incremental Training for OpticDisc Segmentation

deep learning·eye fundus image segmentation·multipledataset training·incremental training·glaucoma

The 27th International Conference on Systems, Signals and Image Processing.

IWSSIP 2020

technical paper

21st edition of EANN (International Conference on Engineering Applications of Neural Networks) should be held in Halkidiki, Greece. Due to current Covid-19 situation the conference will be held virtually and Underline will host a repository of video talks and preserve them for the future. Videos will be released on Underline one day after the conference. 

Performance evaluation an image data set on a CNN from scratch and on a pre-trained network.

Factors affecting accuracy of Convolutional Neural Network using VGG-16

Hyperspectral satellite imagery is now widely being used for accurate disaster prediction and terrain feature classification. However, in such classification tasks, most of the present approaches use only the spectral information contained in the images. Therefore, in this paper, we present a novel framework that takes into account both the spectral and spatial information contained in the data for land cover classification. For this purpose, we use the Gaussian Maximum Likelihood (GML) and Convolutional Neural Network (CNN) methods for the pixel-wise spectral classification and then, using segmentation maps generated by the Watershed algorithm, we incorporate the spatial contextual information into our model with a modified majority vote technique. The experimental analyses on two benchmark datasets demonstrate that our proposed methodology performs better than the earlier approaches by achieving an accuracy of 99.52% and 98.31% on the Pavia University and the Indian Pines datasets respectively. Additionally, our GML based approach, a non-deep learning algorithm, shows comparable performance to the state-of-the-art deep learning techniques, which indicates the importance of the proposed approach for performing a computationally efficient classification of hyperspectral imagery.

A Novel Spatial-Spectral Framework for the Classification of Hyperspectral Satellite Imagery

Exercise Recognition is relevant in many high impact domains, from healthcare to recreational activities to sports sciences, which faces many challenges when deployed in the real world. 
For instance, typical lab performances of Machine Learning~(ML) models, are hard to replicate, due to differences in personal nuances, traits and ambulatory rhythms.
Thus effective transferability of a trained model depends on its ability to adapt and personalise to a new user or a user group.
We look at person-agnostic and person-aware methods of evaluation to identify transferability of generic deep learning models and introduce a personalised algorithm to improve performance in different user groups. 
Our findings show that Exercise Recognition, when compared to results with other HAR tasks, to be a far more challenging personalisation problem.

Transferability of Personalised Exercise Recognition Models

In this presentation I introduce my paper Instance Weighted Clustering: Local Outlier Factor and K-Means. I briefly explain the area, inspirations, methodology and findings. Instance Weighted clustering is easy to understand and produces robust clustering results, the technique can be applied to variety of k-means like algorithms.

Instance Weighting Local Outlier Factor based Approach

There has been increasing popularity in medical text mining due to its vast applications in the field of disease prediction and clinical Recommendation systems. Radiology reports possess rich information depicting radiologists investigations on the health conditions of the patients in associated radiology images. However, radiology reports exist in a free-text unstructured format consisting of valuable information for disease prediction. This information cannot be easily retrieved and utilized for prediction without suitable text mining techniques. The medical dataset available in the current procedure is small, domain-specific and restricted to the institution. However, data is one of the critical factors to power Machine Learning (ML) and Deep Learning (DL) models. To overcome the above challenge of predicting disease in the low data condition, we present a practical Deep Learning framework that combines a Knowledge Base (KB) with the Deep Learning for accurate text mining and predicting the lung diseases from the unstructured radiology free-text reports. We adopt Glove word embeddings with the KB trained on large corpus for effective text modelling. Further, we incorporate Convolutional Neural Network-based Discriminative Dimensionality Reduction (CNN-DDR) to obtain the most discriminative feature vector. Finally, a fully connected Deep Neural Network (DNN) is leveraged as the prediction model to detect the diseases. We applied the proposed framework to predict the lung diseases on radiology reports from both publicly available Indiana University (IU) dataset and data collected from the private hospital. We benchmark the performance of the proposed framework, which outperforms against the standard ML Techniques.

Medical Knowledge-based Deep Learning Framework for Disease Prediction on Unstructured Radiology Free-Text Reports under Low Data condition

In this study, artificial bee colony-based feature selection methods are employed in order to overcome above mentioned problems which is observed in Parkinson’s Disease classification datasets. The fitness values were computed using the k-NN and the algorithms were tested on a public data set. At the end of the study, it is observed that binary versions of artificial bee colony quite successful in feature selection and improves the results in the current literature.

Feature selection with artificial bee colony algorithms for classifying Parkinson’s Diseases

The main purpose of this paper is to find the appropriate fuzzy implication
among the parametric and the three known (Kleen-Dienes, Lukasiewicz
and Reichenbach) from the literature implications which will provide the
smallest square error if adapted to our climatic data.

Application of Algorithmic Fuzzy Implications on Climatic Data

Seismically generated landslides represent one of the most damaging hazards associated with earthquakes in countries with high seismicity. The delineation of prone to coseismic landsliding areas is crucial in order to predict the occurrence of earthquake-induced landslides and consequently reduce the relevant risk. The goal of this study is to investigate the correlation of the pattern of coseismic landslides with geological and topographical variables with the volume of landslides based on fuzzy logic and machine learning techniques. For this task, a real dataset of 421 and 767 instances for years 2003 and 2015 respectively from the island of Lefkada was used. A new approach based on Fuzzy C-Means Algorithm and Ensemble Subspace k- Nearest-Neighbors (Ensemble Subspace k-NN) is proposed. Landslides were classified according to their severity with a success rate of 99.5% and 98.7% for 2003 and 2015 respec-tively. The performance of the proposed approach was evaluated using “One Versus All” Strategy, calculating Accuracy, Sensitivity, Specificity, Precision and F-1 Score for each cluster.

Classification of Coseismic Landslides Using Fuzzy and Machine Learning Techniques

Human activity recognition and analysis has always been one of the most active areas of pattern recognition and machine intelligence, with applications in various fields, including but not limited to exertion games, surveillance, sports analytics and healthcare. Especially in Human-Robot Interaction, human activity understanding plays a crucial role as household robotic assistants are a trend of the near future. However, state-of-the-art infrastructures that can support complex machine intelligence tasks are not always available, and may not be for the average consumer, as robotic hardware is expensive. In this paper we propose a novel action sequence encoding scheme which efficiently transforms spatio-temporal action sequences into compact representations, using Mahalanobis distance-based shape features and the Radon transform. This representation can be used as input for a lightweight convolutional neural network. Experiments show that the proposed pipeline, when based on state-of-the-art human pose estimation techniques, can provide a robust end-to-end online action recognition scheme, deployable on hardware lacking extreme computing capabilities.

A compact sequence encoding scheme for online human activity recognition in HRI applications

Recent advances in optical sensor technologies and those of Geoinformatics can now support very large scale high definition multispectral and panchromatic images. This capability allows the use of remote sensing for the observation of complex ecosystems on earth. It supports sustaining biodiversity, precision agriculture, the management of land, crops, and parasites. It also supports advanced quantitative studies of biophysical and biogeochemical cycles in coastal or inland waters. One of the key applications that will allow the development of new types of decision support systems, is that of precise and effective scene classification. This will offer considerable advantages to business, science, and engineering. This work proposes a novel and very effective system based on geographic object-based scene classification in remote sensing images. More specifically the proposal includes a deep systems architecture of the Residual Neural Network (ResNet) type. This, omitting some layers in the early stages of training, achieves an effective simplification of the network by eliminating the Vanishing Gradient Problem (VGP) which troubles other deep learning architectures. Furthermore, the proposed system uses the softmax activation instead of the sigmoid function in the last level and the fully residual network was trained using the novel AdaBound algorithm that employs dynamic bounds on their learning rates. This achieves a smooth transition to stochastic gradient descent, tackling the noise dispersed points of misclassification with great precision. This is something that other spectral classiﬁcation methods cannot handle. The proposed system has been tested, very successfully, in scene identification from remote sensing images. This confirms that it could be further used in advanced level processes for Large-Scale Geospatial Data Analysis such as cross-border classification, the recognition, and monitoring of certain patterns and multisensor data fusion.

Downloads

Next from EANN 2020

Factors affecting accuracy of Convolutional Neural Network using VGG-16

Similar lecture

Mask Overlaying: A Deep Learning Approach for Individual Optic Cup Segmentation from Fundus Image

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from EANN 2020

Factors affecting accuracy of Convolutional Neural Network using VGG-16

Similar lecture

Mask Overlaying: A Deep Learning Approach for Individual Optic Cup Segmentation from Fundus Image

Downloads