Abstract:

To tackle difficulties of the segmentation of the interested region in complex background, a deep-learning segmentation algorithm based on the fully convolutional network and the dense conditional random field is proposed. First, the fully convolutional network is leveraged for pixel-level feature extraction to obtain the coarse segmentation result. Then, the dense conditional random field which is used to optimize the context information is performed for detailed segmentation. Five-fold cross-validation experiments are carried out on an actually acquired infrared thermal image of the solar panel. Experimental results show that the proposed algorithm has an average precision rate of 89.96%, an average recall rate of 94.55%, an average F₁ index of 0.9118 and an average J index of 0.8687. At the same time, the algorithm achieves the best maximum precision rate of 93.35%, a maximum recall rate of 97.59%, a maximum F₁ index of 0.9562 and a maximum J index of 0.9125 compared with those by main existing algorithms. Moreover, this method takes less time and requires less manual interference. In conclusion, the proposed algorithm is capable of the segmentation of the interested region in the infrared thermal image effectively in the complex background.

Key words: infrared thermal image, segmentation, fully convolutional network, dense conditional random field