Characterization and measurement of monocrystalline-silicon dislocation density are the important parameters for detecting the crystal growth quality and studying the dislocation formation mechanism. Based on atypical characteristics of dislocation corrosion pits such as large differences in morphology and complex background, as well as low accuracy and efficiency of traditional artificial optical microscopy detection, an improved YOLOv5 algorithm was proposed to detect the density distribution of dislocation corrosion pits of monocrystalline silicon. The attention mechanism was introduced based on the original YOLOv5 algorithm to optimize the network structure and strengthen the calculation ability of the model. The network detection accuracy was further improved by strengthening the feature fusion, and the loss function was optimized to enhance the accuracy of positioning and improve the training speed. The experimental results show that the improved algorithm can detect monocrystalline-silicon dislocation pits of different corrosive fluids with accuracy of 93.52% and 98.82%, respectively, the mean average precision (mAP) can reach 96.17%, and the frame rate can reach 47 frame/s, which satisfies the requirements of real-time detection.