The reproducibility of laser powder bed fusion (LPBF) technology in aerospace industry manufacturing is seriously affected by defects, and the defects in the powder spreading process have a significant impact on part quality. In this paper, a detection method based on the real-time semantic segmentation algorithm bilateral segmentation network (BiseNetV2) model and weighted loss is proposed for realizing category identification and location segmentation of powder spreading defects. In addition, a model pruning technique is utilized to optimize the size and performance of the deep learning (DL) model, and the lightweight model is deployed on computers in the monitoring system using the TensorRT technique. The results show that the BiseNetV2 model combined with weighted loss is able to detect five types of powder spreading defects with an average accuracy of 81.23%. The lightweight model obtained by pruning technique significantly reduces the model size by 13.39% while sacrificing 0.44% accuracy. Utilizing the TensorRT technique accelerates the deep learning model inference process and reduces the detection time to 5.94 ms with half-precision floating-point 16 (FP16) data.