Convolutional Neural Networks (CNNs) are a widely accepted means of solving complex classification and detection problems in imaging and speech. However, problem complexity often leads to considerable increases in computation and parameter storage costs. Many successful attempts have been made in effectively reducing these overheads by pruning and compressing large CNNs with only a slight decline in model accuracy. In this study, two pruning methods are implemented and compared on the CIFAR-10 database and an ECG arrhythmia classification task. Each pruning method employs a pruning phase interleaved with a finetuning phase. It is shown that when performing the scale-factor pruning algorithm on ECG, finetuning time can be expedited by 1.4 times over the traditional approach with only 10% of expensive floating-point operations retained, while experiencing no significant impact on accuracy.

This thesis deals with human recognition on a videorecording. Convolution neural network was used for face recognition, from which we will get multidimensional vector, which will allow to determine person’s identity. There are demands imposed on the system, for it to be able to work in real time and could be used for example for person recognition at various conferences, or as a part of security system. Whole system is written in Python language. Part of this thesis is dataset in form of videorecords with persons.

This thesis deals with the issue of detection of anatomical structures in medical images using convolutional neural networks (CNN). At first there are described methods of machine learning, convolutional neural networks and selected methods for detection using CNN. In this work was created a database of annotated CT images of ten anatomical structures (head, heart, aorta, left and right lung, spine, liver, left and right kidney, spleen). A method for detecting these structures was designed, that contains two approaches of region proposals from image, CNN and postprocessing to obtain the detection result. The designed algorithm was implemented in the Python programming language using the TensorFlow library. Obtained results of validation of the network and the detection results are presented and discussed in the last chapter.

This work focuses on automatization of neural network design via the so-called neuroevolution, which employs evolutionary algorithms to construct artificial neural networks or optimise their parameters. The goal of the project is to design and implement an evolutionary algorithm which can be used in the process of designing and optimizing topologies of convolutional neural networks. The effectiveness of the proposed framework was experimentally evaluated on tasks of image classification on datasets MNIST and CIFAR10 and compared with relevant solutions. The results showed that neuroevolution has a potential to successfully find accurate and effective convolutional neural network architectures.

This thesis deals with the detection of intracranial haemorrhages and their type classification in head CT images. The method of haemorrhages detection is based on a series of classifiers of the presence and type of haemorrhages in 2D CT slices in axial, sagittal and coronal plane, that may localise the bleedings and determine their types. The classifiers are based on the convolutional neural network architecture Inception-ResNet-v2. The head CT dataset CQ500 which is made available for public access, is used for the experiments. The thesis describes an additional manual annotation of the data, as the available annotations are insufficient for the purposes of the experiments. This thesis includes a theoretical basis of the essential medical knowledge, machine learning based classification and detection methods, and the detection algorithm proposal, realisation and testing. The algorithm performance is evaluated and discussed together with the potential implementation of the algorithm in computer-aided diagnosis systems.

This thesis deals with the issue of reconstruction of damaged fingerprints using artificial neural networks. At first, the fingerprint structure is analyzed, after that, the methods that can be used to improve fingerprint quality are described. An introduction to neural networks is given for understanding the basics of artificial neural networks. After choosing the right architecture for the neural networks, the process of its learning is described. A simple graphic user interface was created for this application, which is able to reconstruct synthetic fingerprints damaged by various warts. Another neural net can detect the location of wart. Tests have proven an increase in the quality of fingerprint by 43,5 % in the dataset with ten inserted warts on each fingerprint. The matching score was increased by 6,5 % on this particular dataset.

The Master’s thesis deals with the issue of the road detection in the outdoor environment using image processing. It is highly required that the methods selected are robust to sudden light changes within the image and effective in detection of wide variety of road surfaces possibly comprising certain kinds of pollution. Two methods have been used in order to reach the desired goal. The initial method uses standard algorithms of the image processing. Main outcome of this method are highlighted road boundaries. The following methodisbasedonconvolutionalneuralnetworks.Inthiscasewehaveclassificationtask. The result of this method is the estimation of the road direction. In the whole process, severalneuralnetworkstructureshavebeendesigned.Afterthenetworktrainingthemost suitable one was selected. Eventually, the results have been retested using newly created test set. Both of these methods are implemented in programming language Python.

Memory wall is becoming a more and more serious bottleneck of the processing speed of microprocessors. The mismatch between CPUs and memories has been increasing since three decades ago. SRAM was introduced as the bridge between the main memory and the CPU. SRAM is designed to be on the same die with CPU and stores temporary data and instructions that are to be processed by the CPU. Thus, the performance of SRAMs has a direct impact on the performance of CPUs. With the application of mass amount data to be processed nowadays, there is a great need for high-performance CPUs. Three dimensional CPUs and CPUs that are specifically designed for machine learning are gaining popularity. The objective of this work is to design high-performance SRAM for these two emerging applications. Firstly, a novel delay cell based on dummy TSV is proposed to replace traditional delay cells for better timing control. Secondly, a unique SRAM with novel architecture is custom designed for a high-performance machine learning processor. Post-layout simulation shows that the SRAM works well with the processing core and its design is optimized to work well with machine learning processors based on convolutional neural networks. A prototype of the SRAM is also tapped out to further verify our design.

Convolutional neural networks (CNNs) have been extensively used in many aspects, such as face and speech recognition, image searching and classification, and automatic drive. Hence, CNN accelerators have become a trending research. Generally, Graphics processing units (GPUs) are widely applied in CNNaccelerators. However, Field-programmable gate arrays (FPGAs) have higher energy and resource efficiency compared with GPUs, moreover, high-level synthesis tools based on Open Computing Language (OpenCL) can reduce the verification and implementation period for FPGAs. In this project, PipeCNN[1] is implemented on Intel DE10-Standard FPGA. This OpenCL design acceleratesAlexnet through the interaction between Advanced RISC Machine (ARM) and FPGA. Then, PipeCNN optimization based on memory read and convolution is analyzed and discussed.

No description available.