【摘要】 在金相组织的晶粒度自动化评估工作中，对晶粒边界识别的精准与否直接影响着金相组织晶粒度等级的评估准确度；针对钢材金相图像中晶粒边界密集程度高、边缘复杂且晶粒边界识别准确性低的问题，提出一种基于轻量型U-net卷积神经网络的金相图像晶界分割方法，该轻量型网络模型将浅层特征层用跳跃连接的方式拼接在上采样过程中，使网络学习到更多的有效特征信息；减少了网络层数并在特征提取过程中添加了一次卷积过程，减少了网络参数量并提高了对晶界的预测速度和准确率；实验结果表明，该方法在117张金相图像测试集上像素准确率达到93.91%、特异度为96.73%、灵敏度为81.6%;与传统U-net网络相比，像素准确率提高了0.2%,网络参数量相对减少了61.5%;本方法对金相晶界分割具有有效性和优越性。更多还原

【Abstract】 In the automatic evaluation of grain size in metallographic tissue, grain boundary recognition accuracy directly affects assessing accuracy in the grain size grade of metallographic tissue. In view of the problems of dense grain boundaries, complex edges and low accuracy of grain boundary recognition in steel metallographic images, a lightweight U-net convolutional neural network-based grain boundary segmentation method is proposed, which splices the shallow feature layers with the jump connections in the upsampling process, so that the network learns more effective feature information, reduces the number of network layers and adds a single convolutional feature extraction process, reducing the number of network parameters and improving the prediction speed and accuracy of the grain boundaries. Experimental results show that the method achieves a pixel accuracy of 93.91%, a specificity of 96.73% and a sensitivity of 81.6% on a test set of 117 metallographic images. Compared with the conventional U-net network, the pixel accuracy is improved by 0.2%, and the number of network parameters is relatively reduced by 61.5%. The method is effective and superior for the metallographic grain boundary segmentation.更多还原