Image semantic segmentation based on fully convolutional neural network and CRF

Image semantic segmentation is a popular research direction in the computer vision field. Semantic segmentation algorithms based on deep learning outperforms the traditional methods. Fully convolutional neural network (FCN) whose fully connected layers are transformed into convolution layers is a kind of convolutional neural network (CNN). In this paper, FCN is used to operate the image semantic segmentation, which could take input of arbitrary size image and implement end-to-end segmentation task. Due to the limited number of training images, some layers are fine-tuned from AlexNet and the dataset is enlarged by mirroring. The hierarchical feature maps from FCN are combined to improve the segmentation effect. Conditional random fields (CRF) is used on the segmentation result of FCN, which takes into account the positional relationship and color features between any two pixels. Experiments show that our method could refine the segmentation result of FCN, especially using CRF as post-processing.