This study presents a weld detection approach using an improved YOLO deep learning network. The network is a combination of a pre-trained YOLO model and post-convolutional neural network (CNN). Our improved YOLO model has been trained and evaluated on a dataset containing 1691 actual 1G weld images. The results achieve high segmentation and localization performance across all models, with mAP50 scores of 98.30 for segmentation and 97.55 for boundary contour detection. The actual results show that both models demonstrated their strong performance. It is remarkable that the post-CNN YOLOv8 model achieves the fastest inference speed at 300.71 ms per image (10.11% faster), outperforming the standard YOLOv8 by reducing the processing time (40.19% faster) while maintaining comparable accuracy. The addition of CNN layers before and after YOLOv8 did not cause performance degradation but instead provided stability in segmentation accuracy.