The direct prediction model for adiabatic film cooling effectiveness distribution and inverse prediction model for design parameters are studied in this article. Convolutional neural networks (CNNs) are trained on a set of simulated adiabatic film cooling effectiveness contours parameterized by blowing ratio, density ratio, mainstream turbulence intensity, injection angle, and compound angle. The direct model and the inverse model are able to approximate the data in the test set with plausible accuracy. The absolute error of spatial averaged effectiveness no larger than 0.03 could be obtained in the test set by a direct model with time consumption less than 1 ms for a single case. The inverse model is the first model of its kind, which accomplished the inverse mapping from contours to parameters. It has been demonstrated that the concatenation of inverse model with the pretrained direct model, which can be treated as a complex loss function, has preferable approximation performance compared with simple mean squared error (MSE) loss function in the training of the inverse model, thus confirming the necessity of adopting specialized modeling strategies for inverse problems.