Khabarlak K. Image Classification and Segmentation Methods Based on Changeable Convolutional Neural Networks

PhD thesis in specialty 122 Computer Science. – Dnipro University of Technology, Dnipro, 2023. Convolutional neural networks show high quality in solving computer vision tasks. A significant amount of research is devoted to the development of neural networks, that target inference on powerful servers. However, in a number of cases their use is complicated for the following reasons: when the Internet connection is unstable or absent at all, when the user does not agree to share private data from his device, when the data volume is too large to be transferred from all devices to the server, etc. If it is necessary to process images on a mobile or low-power device, a number of problems arise: — such devices have limited computing resources, and the network inference might be unacceptably long for the target task. Therefore, deep convolutional neural network architectures with many parameters that show high quality on servers need to be modified for mobile applications; — inference when running on battery implies that the number of computations should be minimized. Because of this, great interest of the scientific community is devoted to the development of mobile neural network architectures that take into account the mobile device limitations at the design stage. Such architectures require the network configuration to be finalized before the start of the training procedure, as a result long training procedure should be repeated after each network architecture adjustment, which is a problem; — if the application is expected to be installed on an IoT device, this adds another category of devices with less computing power and presents the neural network developer with a choice: either to train one network that will be fast enough for all devices, but potentially have poor performance; or to train a separate network for each category of devices, which, given the long training time of deep neural networks, will significantly increase the cost of development of the system. The purpose of the work is to accelerate convolutional neural network training and inference for the tasks of image classification and segmentation without recognition quality loss (or with as little loss as possible) by developing changeable neural networks and their training methods. By changeable neural network we mean a convolutional network with changeable complexity. Scientific novelty of the obtained results: — for the first time, the changeable convolutional neural network and its training method were developed for the tasks of classification and segmentation. In contrast to the existing ones, changeable neural networks enable configuration selection among the ones with different computational complexities during or after training. On the ImageNet dataset the developed neural network in terms of efficiency (in the sense of recognition quality/execution time ratio) took fifth place among the 17 considered state-of-the-art neural network architectures on the ImageNet dataset. On the CamVid dataset the speed up is above 6 % without quality loss; — for the first time, the Λ-patterns method of optimization meta-learning acceleration was developed, which, unlike existing methods, allows to change neural network complexity during training and, consequently, speed up the neural network few-shot adaptation by 7.5 % with the quality loss below 0.4 %.