A significant challenge in the optical microscopy analysis of weld microstructures (weld metal or HAZ) is that they involve multiple different types of features (e.g. AF, FS(A), FS(NA), M) that can only be recognized manually. Standard image analysis techniques are based on contrast and threshold; in some cases, they can also recognize edges. However, standard software is unable to distinguish features of similar color and size, but differing in their shape, especially at the relatively low resolution of optical microscopy, and with the limitations and variability of etching techniques. For this reason, when quantitative analysis of microstructures is needed, point counting techniques are used (identifying approximately 1000 points in a grid, taking typically 30-60 minutes for each micrograph). Modern software allows to replace point counting with segmentation techniques, but that segmentation is still performed manually. This presentation will introduce current results in applying modern machine learning techniques to identify features in complex microstructures. The technique is based on convolutional neural networks (CNN), and involve data augmentation. Current results can achieve 90% accuracy using only 10 manually labelled micrographs. An interesting fact discovered is that manual labelling is inconsistent and not universally agreed upon specialists. The machine learning algorithm enabled instantaneous automatic labelling of complex microstructures opening the door for data mining of large amounts of stored micrographs and real-time quality control. The algorithm also draw the attention of features that are typically neglected without discussion. In addition, in contrast with standard techniques, machine learning enables the continuous improvement of the microstructure recognition ability with feedback from the user.