The present article demonstrates an effective approach to model the in-silico recognition of phases in steel microstructures using the machine learning and image processing technologies. It is based on a comprehensive microstructure dataset of 192 typical steel SEM images that was created by in-house experimentation, covering a wide variation across types phases and their volume fractions, grain sizes and magnifications. The selective Gray Level Co-occurrence Matrix (GLCM) feature vector extracted from the Simple Linear Iterative Clustering (SLIC) segmented phase patches was effectively used to create a numerical dataset from the microstructure image dataset. A Multi-Layer Perceptron (MLP) model was trained for automatic and accurate recognition of the constituent phases in the microstructure using this GLCM feature vector data. The trained MLP model was able to achieve a network loss of 0.1156 with a prediction accuracy of 83.92% in training dataset and that of 0.25 and 83.5% in the validation dataset, respectively. The prediction accuracy of the proposed MLP model was found to be adequate with the confidence level in automatic recognition of various phases in the range of 77–100%. The quantitative analysis results of the phase regions and phase boundaries from each microstructure appears to be reasonably accurate. The variation in prediction of phase volume fraction against the magnifications were obtained as 16.90 ± 4.0, 70.6 ± 6.0 and 99.97 ± 0.02 (IF steel microstructure) for pearlite, martensite and ferrite phases respectively. The obtained results have been discussed in context of the knowledge of steel microstructures.