At present, most domestic water plants use empirical methods for coagulant dosing control. In order to realize the intelligent dosing of coagulant in water plants, this research has built an intelligent dosing system based on alum floc image recognition. The system combines the YOLOv5 alum floc recognition algorithm and the Linear Regression dosing decision algorithm. And on this basis, a 7-dimensional fully connected BP neural network was added for training through a sample set of (563, 7) (563 samples containing 7 parameters such as the number of alum flocs, the average equivalent diameter of alum flocs, and the inflow flow rate). The optimal weights for each layer were calculated and determined, resulting in a linear regression model with a minimum loss value of 0.018. The production test showed that the detection accuracy of alum floc target was 83.5%, and the predicted dosage was 11.0% lower than the original empirical value. Compared with the traditional control method, the system has lower time ductility, stronger reliability and lower coagulant consumption, which have effectively reduced the production and management costs of dosing in water plants.