An Image Segmentation Method for the blind sidewalks recognition by using the convolutional neural network U-net

Abstract

The image segmentation for the blind sidewalks recognition is important for the blind guide system, which can help the visually impaired people to go out freely. In the paper, we have proposed an image segmentation method for blind sidewalks recognition by using the convolutional neural network (CNN) U-Net. It is theoretically analyzed and numerically proved that the blind sidewalk region can be effectively segmented without requiring a large number of the labeled training samples by using the proposed method. And the mean value for the dice similarity coefficient (DSC) is more than 90% in the proposed method. Besides, the running time for random image segmentation is less than 0.5 second, which can easily satisfy the request for real-time image recognition.