基于卷积神经网络改进的图像自动分割方法

针对图像序列三维重建中多视角目标分割需要人工参与任务繁重的问题，提出一种基于卷积神经网络改进的图像自动分割方法。首先将序列图像去噪处理、归一化并进行语义标注后制作数据集，然后对改进的融合多尺度特征和残差连接的卷积神经网络进行训练，得到优化后的卷积神经网络分割模型，最后将预分割图像加载到优化的分割模型中得到归一化的掩码图,再利用三次样条插值法将其恢复分辨率后与原图做自定义的掩码操作得到高清分割结果。本文以主流分割软件PhotoShop分割结果为参考标准进行对比，实验结果证明，该方法的准确率与参考标准接近，而且可实现批量自动分割，较好的解决三维重建中目标分割任务繁重的问题。

Improved image automatic segmentation method based on convolution neural network

Aiming to solve the heavy problem of multi-angle target segmentation in 3D reconstruction of image sequence, this paper proposed a automatic image segmentation method based on convolution neural network. Firstly, it denoised normalized and semantically annotated the sequence image in order to produce the data set. Then, it trained the convolution neural networks, which were improved with fusion multi-scale feature and residual connection ,and obtained the optimized convolution neural network segmentation model. Finally, it loaded the pre-segmentation image into the optimized model to get the normalized mask pattern, used the cubic spline interpolation methods to restore the resolution and result of the HD segmentation with the original mask. In this paper the result of the main segmentation software Photoshop segmentation are the reference standard. The experimental results show that the accuracy of the method is close to the reference standard, and the batch segmentation can be realized automatically, which can solve the heavy problem of the target segmentation task in 3D reconstruction.