Detection and Control System for Automotive Products Applications by Artificial Vision Using Deep Learning Abdelhamid El Wahabi 1(&) , Ibrahim Hadj Baraka 1 , Salaheddine Hamdoune 1 , and Karim El Mokhtari 2 1 LIST, UAE Faculty of Science and Technology, Tangier, Morocco elwahabi.abdelhamid@gmail.com, i.baraka@gmail.com, shamdoun@hotmail.com 2 Data Science Laboratory, Ryerson University, Toronto, Canada karim@elmokhtari.com Abstract. Object recognition is among the most important subjects in computer vision, it has undergone a huge evolution during these last decades, but in the last years artificial intelligence has seen the appearance of Deep Learning, and through the efforts of researchers, Deep Learning is having great success, its applications have touched on different fields, such as robotics, industry, auto- motive … In this context, in collaboration with an Automotive components manufac- turer and FST faculty of sciences and technologies of tangier (UAE University) have taken the initiative to develop and implement an object recognition and inspection system for automotive products application that requires a good accuracy of image classification using the Deep Learning which is the purpose of this paper. This report summarizes the work done within this Company concerning the development and implementation of a system aims to realize an artificial vision system for the inspection of automotive products based mainly on the “Deep Learning” method. Before, during and after manufacturing, many products in automotive sector (electronic components, …) are subjected to a visual inspection phase, in this context we have replace this phase by our vision system so that the piece will be accepted or not accepted, as well as to act to parameters (for example: winding shape, welding quality …) in the case of not accepted. The convolutional neural networks have become advanced methods for classification and detection of objects over the last five years. At present, they work better than conventional image processing method set, on many image classification data sets. Most of these datasets are based on the notion of concrete classes. In this paper, we present a new set of image classification data as well as object detection data, which should be easy for humans to solve, but its varia- tions are dif ficult for CNN. The classification performance of popular CNN architectures is evaluated on this dataset and variations of this dataset may be of interest for future research. Keywords: Dataset Á AI Á CNN Á ML Á ReLu © Springer Nature Switzerland AG 2020 M. Ezziyyani (Ed.): AI2SD 2019, AISC 1104, pp. 224–241, 2020. https://doi.org/10.1007/978-3-030-36671-1_20