|
|
Known-Item Search in Image Datasets Using Automatically Detected Keywords
Souček, Tomáš ; Lokoč, Jakub (advisor) ; Peška, Ladislav (referee)
Known-item search represents a scenario, where a user searches for one particular image in a given collection but does not know where it is located. The thesis focuses on the design and evaluation of a keyword retrieval model for known-item search in image collections. We use a deep neural network trained on a custom dataset to annotate the images. We design complex yet easy-to-use query interface for fast image retrieval. We use/design several types of artificial users to estimate the model's performance in an interactive setting. We also discuss our successful participation at two international competitions. 1
|
|
|
Environment for Lifting
Kubový, Jan ; Pelikán, Josef (advisor) ; Lokoč, Jakub (referee)
The aim of the thesis is to create a library that will provide ease way to cre- ating and experimenting with computing networks. The concpet of computing netowork can be explained as algorithms whitch can be devided into small simple parts (nodes). The main focus of this library is to easily experiment with trans- formations based on lifting. There are inverse operations for these connections, which are used for lossless compression of data or signal. Emphasis was put on the simplicity of creating new nodes and subsequent connections. An integral part of the work is also an example of several transformations based on lifting.
|
|
|
Visual Question Answering
Hajič, Jakub ; Straka, Milan (advisor) ; Lokoč, Jakub (referee)
Visual Question Answering (VQA) is a recently proposed multimodal task in the general area of machine learning. The input to this task consists of a single image and an associated natural language question, and the output is the answer to that question. In this thesis we propose two incremental modifications to an existing model which won the VQA Challenge in 2016 using multimodal compact bilinear pooling (MCB), a novel way of combining modalities. First, we added the language attention mechanism, and on top of that we introduce an image attention mechanism focusing on objects detected in the image ("region attention"). We also experiment with ways of combining these in a single end- to-end model. The thesis describes the MCB model and our extensions and their two different implementations, and evaluates them on the original VQA challenge dataset for direct comparison with the original work. 1
|
|
|
Comparison of signature-based and semantic similarity models
Kovalčík, Gregor ; Lokoč, Jakub (advisor) ; Mráz, František (referee)
Content-based image retrieval and similarity search has been investigated for several decades with many different approaches proposed. This thesis fo- cuses on a comparison of two orthogonal similarity models on two different im- age retrieval tasks. More specifically, traditional image representation models based on feature signatures are compared with models based on state-of-the-art deep convolutional neural networks. Query-by-example benchmarking and tar- get browsing tasks were selected for the comparison. In a thorough experimental evaluation, we confirm that models based on deep convolutional neural networks outperform the traditional models. However, in the target browsing scenario, we show that the traditional models could still represent an effective option. We have also implemented a feature signature extractor into the OpenCV library in order to make the source codes available for the image retrieval and computer vision community. 1
|
|
|
Effective visualization for interactive video exploration
Pavlovský, Jan ; Lokoč, Jakub (advisor) ; Grošup, Tomáš (referee)
In this thesis we introduce an innovative approach to visualisation and search results presentation for large video collection search and browsing. The general problem of video search is analysed and discussed in comparison with other current software tools and methods used for video search. A specific visualisa- tion method and algorithm for its generation is then proposed and discussed. We evaluated the methods both, empirically and by a user study. Based on the results, we chose the best possible algorithm settings for interactive video search and applied them. A simple experimental software tool implementing the proposed methods is developed focusing on the visualisation components. 1
|
|
|
Aplikace umělých neuronových sítí pro detekci malware v HTTPS komunikaci
Bodnár, Jan ; Lokoč, Jakub (advisor) ; Somol, Petr (referee)
A huge proportion of modern malicious software uses Internet connec- tions. Therefore, it is possible to detect infected computers by inspecting network activity. Since attackers hide the content of communication by com- municating over encrypted protocols such as HTTPS, communication must be analysed purely on the basis of metadata. Cisco provided us a dataset containing aggregated metadata with additional information as to whether or not each sample contains malicious communication. This work trains neu- ral networks to distinguish between infected and benign samples, comparing different architectures of neural networks and providing a comparison with results achieved by different machine learning methods tried by colleagues. It also seeks to create a mapping which maps samples of communication into a space where different samples of malicious communication created by a sin- gle malware family form clusters. This may make it easier to find different computers infected by a virus with known behaviour, even when the virus cannot be detected by the detection system. 1
|
|
|
Efficient kNN classification of malware from HTTPS data
Maroušek, Jakub ; Lokoč, Jakub (advisor) ; Galamboš, Leo (referee)
An important task of Network Intrusion Detection Systems (NIDS) is to detect malign com- munication in a computer network traffic. The traditional detection approaches which analyze the content of network packets, are becoming insufficient with an increased usage of encrypted HTTPS protocol. The previous research shows, however, that the high-level properties of HTTPS commu- nication such as the duration of a request or the number of bytes sent/received from the client to the server may be successfully used to detect behavioral patterns of malware activity. We study approximate k-NN similarity joins as one of the methods to build a classifier recognizing malign communication. Three MapReduce-based and one centralized approximate k-NN join methods are reimplemented in order to support large volumes of high-dimensional data. Finally, we thoroughly evaluate all methods on different datasets containing vectors up to 1000 dimensions and compare multiple aspects concerning scalability, approximation precision and classification precision of each approach.
|
|
|
Using Metric Indexes For Effective and Efficient Multimedia Exploration
Čech, Přemysl ; Lokoč, Jakub (advisor)
The exponential growth of multimedia data challenges effectiveness and efficiency of the state-of-the- art retrieval techniques. In this thesis, we focus on browsing of large datasets using exploration approaches where query cannot be particularly expressed or where the need of some general notion of the dataset is required. More specifically, we study exploration scenarios utilizing the metric access method PM-Tree which natively creates a hierarchy of nested metric regions. We enhance the PM- Tree for the exploration purposes and define different traversing and querying strategies. Further, we investigate range multi-query (range query defined by multiple objects) approaches. We propose new effective and efficient cut-region range query and compare the new query with other approaches for efficient multi-query processing. Finally, we implemented all new methods and strategies for the PM- Tree into multimedia exploration framework and tested browsing algorithms on a game-like demo application.
|
|
|
Object recognition using 3D convolutional neural networks
Moravec, Jaroslav ; Lokoč, Jakub (advisor) ; Straka, Milan (referee)
Title: Object recognition using 3D convolutional neural networks Author: Jaroslav Moravec Department: Department of Software Engineering Supervisor: RNDr. Jakub Lokoč, Ph.D., Department of Software Engineering Abstract: With the fast development of laser and sensor technologies, it has become easy to scan a real-world object and save it in a digital format into a persistent database. With the rising number of scanned 3D objects, data man- agement and retrieval methods become necessary. For various retrieval tasks, effective retrieval models are required. In our work, we focus on effective classifi- cation and similarity search. The investigated approach is based on convolutional neural networks representing a machine learning method that boomed in recent years. We have designed and trained several architectures of 3D convolutional neural networks and tested them on state-of-the-art benchmark 3D datasets for 3D object recognition and retrieval. We were also able to show that the trained features on one dataset can be then used to predict class labels on another 3D dataset. Keywords: Object recognition, 3D convolution, neural networks
|
|
|
Detection of malignant melanoma in histological sample using deep neural networks
Frey, Adam ; Lokoč, Jakub (advisor) ; Straka, Milan (referee)
The aim of this thesis is to create a classification method for detection of ma- lignant melanoma in high-resolution digital images. Deep convolutional neural networks were used for this task. At first, a short overview of malignant melanoma and ways to detect it is presented. Deep convolutional neural networks are also introduced with a special attention given to models used further in this work. Several ways to generate samples from the provided histological images are discussed, and several experiments are evaluated to decide how to maximize the accuracy of employed classification methods. The thesis then focuses on several neural network structures used for image classification and their possible utiliza- tion for the given task. The emphasis is laid on the transfer learning, a method used for modifying already trained models for different tasks. This method is then used for training several classifiers. Further on, several methods for the visualization of model results are discussed with some of them implemented. The experiments show promising results on par with other studies dealing with similar problems. Several possibilities for further development are listed in the conclusion.
|
guest ::
