PARALLELIZATION ALGORITHM OF IMAGE CLASSIFICATION ACCORDING TO “SMART HOUSE SECURITY”

Abstract

Many literature sources were analyzed regarding the practicality of choosing the CNN model. They described all the strengths of this neural network structure. The problem of classification using the neural model CNN and OpenCV is considered. With the help of OpenCV, you can process video streams in real-time, and already with the help of a trained model, you can classify interlaced images. Thus, the study aims to model the architecture of the model in such a way as to minimize the classification time and increase its accuracy. You can also improve classification speed using multithreading. For a complete analysis of the study, the working principle of CNN should be considered. It is a class of deep neural convolutional networks that can recognize and classify certain image features and is widely used for visual image analysis. During the experiments, the CNN classification model is used. To get information about the most optimal options, various parameters are selected to compare the results, such as optimizers and the number of epochs. During the experiments, training took place with different sizes of epochs and optimizers. To improve the speed, the two most effective ones were selected, namely “Adam” and “RMSProp”, because when choosing the others, the accuracy dropped sharply to values less than 40%. To determine the effectiveness of one over the other, we performed training and classification of images with the same parameters. When the number of epochs increases, it is possible to notice positive dynamics in terms of accuracy and negative dynamics in terms of losses on both models when the number of epochs increases. This confirms the effectiveness of CNN for its direct purpose, namely image classification. Experiments on image size were also carried out in work. As a result of this experiment, the training time was reduced by another 8 s per epoch because the input image pixel matrix size became smaller. Also, the conducted experiments proved that with the optimal image size selection, the processing speed could also be increased. After selecting all the optimal input parameters of the neural network, the parallelization process of the classification was analyzed. The work offers ways to optimize the model.