A lightweight object detection network YOLO-SEC based on YOLOv5s was proposed to resolve the problem that the arc shape in ultra-narrow gap welding process was difficult to detect online and describe quantitatively. Three lightweight convolutional modules, namely AD-Shufflenet, LC3 and Conv-Maxpool, were designed and applied to the backbone network and neck network of baseline network. Single detection layer inference structure was used in the detection head network, so that the network complexity was effectively reduced. The k-means+ algorithm was designed to cluster labeled anchor frames in the data set and obtain the preset anchor frame size required for network training, and the EIoU positioning loss function was introduced to improve the accuracy of the network. The monocular linear camera calibration algorithm was introduced into the prediction frame drawing function, so that the pixel length could be converted into the actual length by calibration before welding, and the quantitative description of arc size was realized. Experimental results showed that the detection accuracy of YOLO-SEC was 99.2%, the measurement error was less than 0.5%, the reasoning speed was 3.1 ms, and the network volume was 0.6 MB. All the above indexes were better than the YOLOv5s network and the same type of lightweight network.