改进U-Net网络的多视觉图像特征张量分割仿真

Simulation of Multi Vision Image Feature Tensor Segmentation Based on Improved U-Net Network

针对图像分割计算量大、噪声因素影响等问题,提出改进U-Net网络的多视觉特征图像分割方法。对同一窗口中的灰度值排序,计算像素点极大值与极小值,根据角度与像素点的关系,检测噪声点,将被污染的噪声点放入集合中,使用其它像素点替换该点,完成滤波;分别从颜色、纹理与形状三个方面提取图像的多视觉特征,为图像分割提供参考依据;利用编码器、解码器和跳跃连接层建立U-Net网络,将提取的特征作为网络输入,新增深度残差模块,经过残差学习,实现特征映射;引入注意力模块,减少特征维度,确定张量权重,利用池化层拼接特征维度,输出最终分割特征张量。实验结果表明,所提方法对于分割目标的敏感度较高,不容易出现过分割与欠分割现象。

This paper put forward a method of segmenting multi-vision image features based on improved U-Net network.At first,we sorted the gray values in the same window,and then calculated the maximum and minimum values of a pixel point.According to the relationship between angle and pixel,we detected noise points and put the contaminated noise points into a set.Moreover,we replaced the point with other pixel points,thus completing the filtering.Furthermore,we extracted multi-vision features of the image from color,texture and shape,and thus providing a reference basis for image segmentation.Meanwhile,we constructed a U-Net network including encoder,decoder and jump connection layer.After that,we used the extracted features as network input,and added a deep residual module to the U-Net network.After residual learning,the feature mapping was achieved.In addition,we introduced the attention module to reduce the feature dimension,thus determining the tensor weight.Finally,we used spliced feature dimensions by pooling layer,thus outputting the segmented feature tensor.Experimental results show that the proposed method is sensitive to the segmentation of the target and is not prone to over-segmentation and under-segmentation.