Návrh a optimalizace obrazových klasifikátorů

Design and optimisation of image classifiers

Object detection is a very important operation in image processing systems such as various surveillance and security systems. This operation is very computationally intensive and it consumes a large amount of resources. The detection can be performed by image classifiers. The development of new image sensors, which have a big resolution and data rate, brings higher requirements on computation resources and also the classification has to be very fast, precise and effective. For these reasons the thesis is focused on image classifier runtime optimization, creation and utilization of the FPGA technology for their implementation. The thesis utilizes classifiers based on AdaBoost, which is a universal classification method, by which whereby various objects can be detected. The AdaBoost belongs to a group of boosting methods, which are based on statistical combination of many weak classifiers to obtain the final result of classification. In this work simple binary image operators - Local Binary Pattern are utilized to create the weak classifiers. These operators have good discriminative capabilities and they are suitable for FPGA implementation. One goal of this thesis is to prove that using application specific classifiers can lead to better results of computation time and effectivity, than results achieved by general classifiers. The accomplishment of this goal is divided into a few parts of which the first is adjustment of the image operator shape for the current task. To solve this issue, genetic algorithm was chosen. The new application specific shapes of operators were designed by this approach and the total classification accuracy was improved by 4 %. The main advantage of the newly designed features is their optimization of classification accuracy of the current application task. The main contribution of this thesis is in introducing a Real-Time Multi-Parallel Object Detector - RT-MPOD. It is based on multiple applications of the AdaBoost - AC evaluation unit, which utilizes an unordered evaluation of weak classifiers to achieve completely parallel processing. The AC unit is designed to evaluate one detection window in each clock cycle. The RT-MPOD reaches very high processing speed - one image pixel per clock cycle and it can process an image stream of up to 300 Megapixels per second (it is for example a video with a resolution of 3000 x 2000 pixels and 50 frames per second), which is nearly a magnitude higher than other introduced architectures. The high processing speed is reached by a few methods, which were used in this work. The first one is utilization of multiple parallelisms during the data processing. The second one is utilization of application specific implementation of each used AC unit. Each classifier is maximally adapted to the current task. The introduced AC unit and also the whole RT-MPOD architecture are specific by a stream character of data processing. It means that there is no need to return to already processed pixels and there is no need of additional buffers. Because each part of the classifier is strongly application specific, the method of automatized classifier design was introduced to simplify the classifier creation. After that the designer works with a tool on a high level of abstraction and he determines the parameters such as accuracy of the classifier, FPGA requirements, training data, throughput and so on. The RT-MPOD is also designed to be easily divided into several computational units in order to be used as a pre-processing unit to pre-select objects of interest.

AdaBoost, Real-Time, classification, RT-MPOD, Local Binary Pattern, genetic algorithm.

KADLČEK, Filip. Návrh a optimalizace obrazových klasifikátorů. Brno, 2016. Ph.D. Thesis. Brno University of Technology, Faculty of Information Technology. 2017-04-28. Supervised by Fučík Otto. Available from: https://www.fit.vut.cz/study/phd-thesis/565/