Laser based directed energy deposition (L-DED) is a promising type of additive manufacturing technology. The non-destructive testing technology for the quality monitoring of L-DED processed parts is becoming more and more demanding in terms of accuracy, real-time, and ease of operation. This paper introduces a new image recognition system based on a deep convolutional neural network, which uses multiple lightweight architectures to reduce detection time. In order to eliminate the interference better, it improves the penalty function, which effectively improves the accuracy. Judging from the detection results of the data set, the accuracy of the model training reaches 94.71%, which achieves a very good image segmentation effect and solves the technical problem of in-situ monitoring of the L-DED process. This system realizes the positioning of the spatters for the first time, and at the same time, the number of spatters and area of molten pool are correlated to the laser scanning speed and the laser power.

基于激光的定向能量沉积 (L-DED) 是一种很有前途的增材制造技术。用于L-DED加工件质量监控的无损检测技术在准确性、实时性和易操作性方面的要求越来越高。本文介绍了一种基于深度卷积神经网络的新型图像识别系统，该系统使用多个轻量级架构来减少检测时间。为了更好地消除干扰，改进了惩罚函数，有效提高了准确率。从数据集的检测结果来看，模型训练的准确率达到了94.71%，达到了非常好的图像分割效果，解决了L-DED过程原位监测的技术难题。