To address the problems of poor real-time performance and insufficient small target detection capability of existing methods for quartz crucible bubble detection, a modified YOLOv5 algorithm for quartz crucible bubble detection, YOLOv5-QCB, is proposed. Firstly, a self-built quartz crucible bubble dataset is constructed, and based on the characteristics of small bubble size and dense distribution, the depth of network down-sampling is reduced to retain rich detailed feature information; meanwhile, the neck using dilated convolution to increase the feature map perceptual filed to achieve global semantic feature extraction; finally, the effective channel attention mechanism is added before the detection layer to enhance the expression of important channel features. The results show that compared with original model, the improved YOLOv5-QCB can effectively reduce the missed detection rate of small bubbles, improve the average accuracy from 96.27% to 98.76%, and reduce the weight by one-half, which can achieve fast and accurate detection of quartz crucible bubble targets.