Deep Learning Solution for Analyzing Visual Imagery in Industrial Applications

Machine learning is one of the hottest topics of the last years in the computer industry. The growing interest on methods for processing large amount of heterogeneous data and new cognitive systems is creating new challenges and opportunities. In the ICT domain, a major effort is spent on improving and applying machine learning, deep learning and in general artificial intelligence techniques. Applications can be seen in various fields, from civil to military, through industrial. This work of thesis is focused precisely on this last area of application and precisely on the image recognition problem, which is addressed using deep learning (DL) models based on a state-of-the-art Convolutional Neural Network. Image recognition is used in the industrial area for the quality control of the products, for tracking, counting and measuring objects, etc. Although high performance devices are often required to perform image recognition operations, one of the most popular market trends is to try to use devices that require a lower amount of electric power to work. Starting from this statement, the goal of this thesis was to try to use the Parallella board, that is a modern low-power parallel general-purpose device, to run two different deep learning models based on Darknet, that is an open source neural network framework. The limitations of the device used, in terms of both performances and available resources, have represented the main challenges of this research and also the starting point for all the solutions found. Different approaches have been followed and investigated to optimize the evaluation times for a single image, with the aim of making the solution found as usable as possible in a real context: from a basic approach that tries to make the most of the board's multicore architecture; to an ad-hoc implementation developed to bypass the main bottlenecks of the device (like the poor amount of memory available) trying to exploit the many cores present in a smart way; finally trying to further improve the performances by intervening directly on the algorithm used for the convolution operation. The different approaches used have been tested and evaluated, allowing to express some considerations on the use of low-power devices for machine learning applications and in particular on the direction that the scientific research could take in order to improve such use.