Browsing by Author "Vrejoiu, Mihnea Horia"
Now showing 1 - 5 of 5
Results Per Page
- ItemConvolutional Neural Networks, Big Data and Deep Learning in Automatic Image Analysis(ICI Publishing House, 2019-03-31) Vrejoiu, Mihnea HoriaIn recent years, there is an increasing amount of talk about artificial intelligence. What actually stands behind artificial intelligence today can be briefly summarized by the syntagm „artificial neural networks“, to which the adjective „deep“ has recently been added. Applications based on these have come to equate and even surpass human performance in many areas. One of the first fields in which they have been developed and in which they have gained a wide spread is artificial vision, respectively image recognition / classification. Without claiming to completely cover the subject, in this paper we propose a review, trying to capture as much intuitively as possible some essential elements and milestones of the history and evolution of artificial neural networks, with the new perspective offered in the last period by the availability of massive data (Big Data) used in conjunction with them as a major complementary, synergistic and convergent factor along with the quality and performance of the deep learning algorithms involved. Also, we analyze the elements and mechanisms that define and compose the convolutional networks in general, their functioning and their specificity with application in artificial vision, as well as two of the first such reference architectures, AlexNet and VGGNet, with their peculiarities and techniques used in the training, validation and testing processes.
- ItemDeep Learning pentru descrierea automată a imaginilor în limbaj natural – Image Captioning(ICI, 2020-01-31) Hotăran, Anca Mihaela; Vrejoiu, Mihnea HoriaImage Captioning (IC) in Computer Vision context refers to the automatic generation of textual descriptions associated with digital images. It is not only the recognition of the objects in these images, but also the description of their properties, as well as the relationships and interactions between them, all expressed textually in natural language, syntactically and semantically correct. Synthetically, the main steps in the automatic generation of textual descriptions associated with the images are: a) – extracting the visual information from the image, and, b) – “translating” it into an adequate and meaningful text. The spectacular developments in the field of deep neural networks and Deep Learning in recent years have led to absolutely remarkable progress also in the field of IC, the quality of the generated descriptive texts being substantially improved. Convolutional Neural Networks (CNN) have been naturally used to obtain essentialized vectorial representations of the image features, and Recurrent Neural Networks (RNN), in particular Long Short-Term Memory (LSTM), were used to decode these representations into phrases in natural language. In this paper we present an overview of the new techniques and methods based on Deep Learning used in the IC field, while also detailing and analyzing, as a case study, one of the best performing ones, using an encoderdecoder architecture combined with a mechanism for focusing the visual attention on the appropriate relevant regions of the image when generating each new word in the output sequence.
- ItemDeep Reinforcement Learning. Studiu de caz: Deep Q-Network(ICI Publishing House, 2019-09-30) Vrejoiu, Mihnea HoriaArtificial Intelligence (AI) became today perhaps the most up-to-date topic in many areas. One of the main goals of AI is to create completely autonomous agents able to interact with the surrounding world and learn by trial and error optimal behaviors in different contexts, perfectible in time. Among the machine learning methods of AI, reinforcement learning (RL) by repetitive interactions with the environment while targeting a purpose plays a particularly important role, besides supervised and unsupervised learning. However, classical RL methods have important limitations in scalability to higher-dimensionality problems. In recent years, supervised and unsupervised learning technologies based on deep learning, using deep neural networks with remarkable properties of approximating complex functions on multi-dimensional spaces, as well as the learning of characteristic hierarchical representations automatically extracted directly from data, with significant dimensional reduction, have had an explosive development, producing astonishing results comparable with, or even surpasing human performance in areas such as object / image recognition, speech recognition, automatic translation etc. The combination of RL with deep learning methods has led to what is now called deep reinforcement learning (DRL), providing new possibilities for producing autonomous agents in multidimensional spaces. This paper is proposing a brief presentation of the DRL field, while also studying and analyzing in detail one of the first successful DRL methods, namely Deep Q-Network developed by Google DeepMind.
- ItemLicense Plate Segmentation in Images Based on per-Block Contrast Analysis and CCA, in(ICI, 2020-06-30) Vrejoiu, Mihnea HoriaALPR based applications are more and more used today. Besides the OCR part, the vehicle registration plate detection in real world images represents the main challenge in LPR. This paper presents a simple, yet quite general, fast and effective method for license plate (LP) segmentation. It is based on the evaluation of a local contrast (high gradient) measure at the level of image blocks, binarization of the downscaled contrast map obtained with these values, and analysis of connectivity between its runs, requiring modest CPU and memory resources. It provides as output not only the locations of detected LPs, but also associated bitmaps, black on white, containing only their constituent alphanumeric characters, aligned horizontally, with no slope, slant or tilt, and free of other parasitic noise. Such black on white bitmaps are directly suitable for further OCR, the correctness and completeness of final LPR strongly depending on the quality of the bitmap provided. Extended experiments carried out on own image set, as well as on other (public) data sets, showed good performance and results of the implemented method in the vast majority of situations, even on certain difficult, poor quality images. Comparison with state-of-the-art (based on deep neural networks, and high-end GPU parallel computing), also proved average good performance on public data sets complying with the minimal requirements of our method.
- ItemNeural Networks and Deep Learning in Cyber Security(ICI Publishing House, 2019-06-30) Vrejoiu, Mihnea HoriaIn the last years, the deep learning (DL) technology using various deep neural network models / architectures became the state-of-the-art in Machine Learning (ML) and Artificial Intelligence (AI), its applications reaching better performances than humans in more and more domains. While traditional ML techniques were mainly based on certain mandatory initial “hand-crafted” feature extraction and engineering phase, the new DL approach is automatically performing this step of specific feature representations extraction directly from the raw input training samples. This intrinsic ability makes it applicable to various issues that cyber security is currently dealing with, such as: intrusion detection, malware classification and detection, spam and phishing detection and binary analysis. In this paper we are intending a brief overview of artificial neural networks and some examples of deep learning based solutions in cyber security.