In recent times significant advancement in quantitative metallography has been taking place in association with the progression of image analysis with advanced machine learning techniques. Consequently, the digital microstructure analysis and automatic phase segmentation emerged as an active research field which is potentially nurturing the structure- properties linkage of materials. However, there are still great challenges in accurately segmenting phases in the complex steel microstructures. To address this problem, a hybrid model based on fully convolutional neural networks (FCNN) and dense conditional random fields (CRF) has been proposed in the present article to quantify and segment the phases accurately and more efficiently with deeper model parameters. In the first step, FCNN has been adopted for feature extraction to perform pixel-wise segmentation irrespective of the input size of micrographic image. Secondly, appearance similarity and smoothness kernel based dense CRF is used for boundary refinement of the segmentation results obtained from FCNN model. The proposed hybrid-model was validated with a variety of complex microstructures which were generated in-house by conducting different heat treatment and some publicly available microstructures. The model was quantitatively compared with some various state-of-the-art techniques to demonstrate the reliability of our model. The results obtained from the proposed model exhibit a reasonable mean intersection over union, f1-score and accuracy for the binary phase as well as for multi-phase microstructures. The final mean intersection over union of our model in each microstructure is over 0.71, f1-score is over 0.81 and pixel accuracy is more than 91 %, the quantitative results depict that our model has better segmentation results in the automatic analysis of the multiphase steel microstructures.