The microstructural characterization of steels is key to optimize their processing route and ultimately better control their properties. It is of particular interest to classify, quantify and localize the different transformation products such as ferrite, pearlite, martensite and the different bainites inside the microstructure. This task requires a considerable amount of time and effort, and it is often performed by different experts, therefore carries a great degree of subjectivity. For years, the metallurgical industry has been looking for alternatives to accomplish this work, trying to minimize time, effort and subjectivity. Artificial intelligence (AI) based techniques are an upcoming and attractive option for materials characterization. Several studies have applied AI to this end on optical and scanning electron micrographs with good results in simple cases but moderate results in more complex cases 1–5. Microstructural images alone may not carry enough information to resolve ambiguous microstructures. Electron Backscattered Diffraction (EBSD) maps allow to access both the microstructural and crystallographic information of the microstructure. This allows one to exploit new features such as orientation relationship between the transformation products, habit planes or spatial distribution of the micro-constituents (packet, blocs, sub-block distribution) to improve the characterization 6–9. Few studies have started to implement machine learning on EBSD-based data for steels’ phase classification 10,11. The question we want to address is: would a neural network succeed in classifying the transformation products with EBSD maps and understand the intricate details that comes along with the orientation data? To answer this question, we propose the discrimination of ferrite and martensite in Dual Phase microstructures as a model study case. The present approach is based on a semantic segmentation strategy through the use of UNET architecture, a deep learning (DL) algorithm in the group of the convolutional neural networks (CNNs). Multiple models have been trained based on two main EBSD-based quantities: Pattern Quality Index and/or raw orientation data in the form of quaternions. Systematic assessment of the architectures hyper-parameters has been performed on a training set in order to optimize the performance. The trained models reached an average accuracy of 94% and above for the considered dataset. Good prediction/classification has been achieved by the models on microstructures with increasing complexity, providing promising results in several cases. Compared to other available approaches in the literature for phase discrimination, the model presented here provides higher accuracies. The proposed technique emerges as a powerful alternative to efficiently classify Dual Phase steel microstructures. Future studies will be dedicated to the extension of this approach to more complex microstructures where multiple transformation products appear.