An Improved U-Net Method for Sequence Images Segmentation

Abstract

In multi-view three-dimensional reconstruction of objects, the accuracy of the image segmentation plays a key role in the accuracy of the model. The traditional Convolutional Neural Network segmentation method often leads to significant feature losses in the target’s edges. It also requires a lot of data for training. Therefore, this paper proposes an improved U-Net method for sequence image segmentation. To begin with, the U-Net structure is used as the basis to solve the problem of feature position information loss and to improve the precision of the edges of segmented objects. Next, multi-scale convolution modules are added on the basis of U-Net structure to increase the network depth and improve feature extraction capability. Then the batch normalization layer is added to solve the problem of vanishing gradient and to accelerate the speed of converged network. Finally, a heat-map channel is added in the input data to prevent errors of classification in similar areas. The experimental results showed that this method ranks higher than the classical U-Net on key indicators, Fl-score and IOU. It can effectively improve the segmentation accuracy, yielding results similar to those of manual segmentation.