Sistema de classificação de imagens utilizando uma rede neural Squeezenet embarcada em uma Raspberry Pi

Abstract

Computer Vision is a field of Artificial Intelligence characterized by the study of existing information in images, identifying their intrinsic characteristics. The study of Computer Vision aims to create artificial models that mimic the analytical skills of human vision, for this, concepts of Digital Image Processing are used to extract information to be studied. Performing these operations requires a large amount of data to be effective and for that, algorithms capable of processing this information are needed. In this context, Deep Learning algorithms are ideal for working with a huge amount of data, as they are efficient and effective. In this way, the use of Neural Networks for this purpose becomes quite suitable, as this tool allows it to be possible to learn from a set of examples so that the generalization of the data is adequate to the examples provided. In the case of images, Convolutional Neural Networks are the state of the art in the area of Computer Vision, and it is possible to observe several applications involving image classification, object identification and face recognition. However, these algorithms are robust and have a complex implementation, having several free parameters that are determined during execution, requiring that the hardware that supports it has high computational capacity to function without errors or with exacerbated execution time. For the case of embedded systems that require low implementation cost, single board computers are commonly adopted, considering that such hardware can be applied in different contexts and have low execution cost. However, these devices are restricted in terms of computational power and a major study of techniques that allow the execution of complex algorithms on their hardware is necessary. Thus, this work aims to present an example of implementation of an image classifier in a Single Board Computer with a Convolutional Neural Network (CNN) running. The concepts of CNNs and Digital Image Processing used during the development of the project are exposed. The developed classifier captures images of handwritten digits and classifies them in real time into 10 classes distributed from 0 to 9. In addition, the developed Digital Image Processing techniques are demonstrated, which use the Gaussian Filter to approximate the images used to the CNN training and the images used during the test of the embedded classifier. The results of the system’s classification prove to be reasonable for the established scenario, being relevant results for the work in question, in particular with regard to the classification accuracy of the system of 76% and a precision of 80% when classifying the images.