A prediction method for iced airfoil aerodynamic characteristics is proposed based on the Convolutional Neural Networks （CNN）. This new method is able to solve the multiple-value problem of the complex ice shape along the wall-normal direction at a specific location of the airfoils. The CNN prediction model can directly predict the lift and drag coefficient at multiple angles of attacks via a rapid process from the iced airfoil image to aerodynamics characteristics. The relative mean errors of the resulting lift coefficient and drag coefficient are smaller than 8%. By comparison， the influence of the convolutional layer number， filter number， and filter size on the model performance is investigated. The features of different layers in the CNN correspond to different filter frequencies， and increasing the number of convolutional layers will capture more features of high frequency. With the increase in the number of filters， more features of the ice shape will be extracted， promoting the model performance. However， more redundant features will appear if the number of filters exceeds a critical value， resulting in degradation of the generalization performance. The required number of filters for drag coefficient prediction is larger than that for lift coefficient， because drag coefficient is affected by both the pressure difference on the airfoil and friction， leading to a larger number of key features required for drag coefficient prediction than that for lift coefficient. Moreover， a larger size of filters will widen the field of view of convolution operation， enhancing the capability of extracting global features， and promoting the generalization performance of the model. The current new method can be applied to real-time prediction and monitor of aerodynamic characteristics of aircraft icing.