During the production of monocrystalline silicon, defects generated during the crystal pulling process are recognized to severely impact product quality. Traditional visual-based defect detection methods, when applied to the detection of small protrusions in crystal pulling images, are confronted with challenges such as slow detection speeds, large parameter volumes, and difficulties in deployment on embedded terminals. In response to these challenges, an improved YOLOv8 object detection model was proposed incorporating a ContextGuided module to enhance the inference efficiency of the model. An efficient DySample was introduced into the feature fusion network to optimize the efficiency and depth of feature fusion. A lightweight network structure was adopted to reduce the complexity and computational demands of the model, making it suitable for devices with limited computing resources. The model has been trained and tested on an industrial dataset, demonstrating a more accurate detection of small protrusions with a mean average precision(mAP) of 97.7%. Compared to YOLOv8n, it exhibits an increase of 11.6% in precision and a reduction in parameter volume by 31.9%, facilitating its deployment on embedded terminals.