基于改进ResNet-UNet的立木图像分割方法

A tree segmentation method based on ResNet-UNet

【目的】针对现有图像处理方法分割立木精度低的问题,提出一种基于改进ResNet-UNet的立木图像分割方法,实现对图像中立木的精确分割。【方法】将拍摄得到的立木图像输入ResNet-UNet深度学习融合网络模型,初步得到较精确的立木分割图;结合自制的后期处理方法对该分割图进行优化处理,准确分割出立木形状。ResNet-UNet模型充分利用了像素之间的语义关联,以ResNet-34残差模块作为ResNet-UNet网络特征提取的基本单位;以U-Net网络的设计思路对图像进行上采样,以实现分辨率还原。去除ResNet-34网络的平均池化和全连接层,改变U-Net网络模型的特征通道数,形成ResNet-UNet网络模型。结合使用Adam一阶优化算法和dice bce loss损失函数实现了立木图像的初步分割。在后期处理阶段设定动态阈值得到前景和背景,避免了使用固定阈值对立木图像的高质量要求。运用强化学习中评分惩奖的思想,对前景和背景分配像素估计值,将该值和模型训练不同次数时的损失率输入惩罚-奖励机制,从而减轻分割结果对预测最终结果的过度依赖,降低网络过拟合对分割精确度的干扰。【结果】经验证,在自然环境、不同光照条件下分割不同品种的立木,平均误分率较传统的ResNet-UNet方法降低了3.5%,假阴率和假阳率较graph cut方法都降低了20%。【结论】使用ResNetUNet方法分割立木具有较高的精确度和较强的鲁棒性。

【Objective】To solve the problem of low accuracy of trees image segmentation using existing image segmentation methods, a new method fused the ResNet-34 and U-Net, so as to achieve accurate segmentation of single target tree in trees image. 【Method】Input the obtained standing wood images into the ResNet-UNet deep learning fusion network model, to obtain more accurate standing wood segmentation map preliminarily. Combined with the self-made post-processing method, the segmentation diagram was optimized to accurately segment the standing wood shape. ResNet-UNet model makes full use of the semantic association between pixels, and takes ResNet-34 residual module as the basic unit for feature extraction of ResNet-UNet network. The image is sampled up with the design idea of U-Net network to restore the resolution. Remove the average pooling and full connection layer of ResNet-34 network, and change the number of characteristic channels of the U-Net network model to form the ResNet-UNet network model. Combined the Adam first-order optimization algorithm and the dice bce loss function to realize the preliminary segmentation of the standing wood image. In the post-processing stage, the dynamic threshold value is set to foreground and background, which avoids the high quality requirement of standing wood image by using fixed threshold value. Based on the idea of scoring and rewarding in reinforcement learning, pixel estimation values are assigned to the foreground and background, and the loss rate of the value and the model is input into the penalty-reward mechanism at different times of training, so as to reduce the over-dependence of segmentation results on the final prediction results and reduce the interference of network overfitting on segmentation accuracy.【Result】It is verified thatunder four different lighting environments, the average false classification rate of different varieties of trees is 3.5% lower than the traditional pure method of Resnet-50 and U-Net, and the false shading rate and false positive rate are 20% lowerthan thegraph cut method.【Conclusion】The experiment proves that this method has high accuracy and strong robustness in the segmentation of single target tree.