Comparison of MS Custom Vision Auto Machine Learning with Algorithms Implementation Methods

This study is devoted to the acquaintance, application, and implementation of computer vision models in the automated Microsoft Custom Vision service, as well as comparing the results of such models with the results of work personally assembled using the PyTorch/Tensorflow frameworks. The base models were represented by convolutional neural networks, which are applied to the image classification task. The study showed that the functionality of the Custom Vision service, although not great, but they represent really strong models, because without searching for the architecture, tuning the hyperparameters of the convolutional neural network, it is not possible to achieve results comparable to automatic machine learning solutions. However, the disadvantages of such models are the complexity of their deployment. As data for the study, we used databases of images of pollen grains of various plants, as well as aviation security images recorded by introscopes at airports. The experiment showed that in terms of precision and recall metrics, the Custom Vision solution provides an advantage of about 2-3% on balanced data and up to 15% on unbalanced ones. Thus, it is shown that using machine learning cloud services, you can get the primary baseline model, which programmers can further develop and improve, and other users can try to launch in production without knowledge of machine and deep learning technologies if the model quality estimation is quite high.