Aiming at the problem that the effective texture features in the space-time image become blurred in complex noisy scenes such as flare, turbulence and rainfall, which limits the accuracy of the existing main direction detection algorithms of texture, a space-time image velocimetry method based on residual network regression model is proposed by combining the idea of deep learning, and the powerful nonlinear learning ability of the model is used to construct a regression prediction function. The residual network regression model is tested by constructing synthetic datasets and natural river datasets containing space-time images of complex scenes. The experimental results show that the detection accuracy of the proposed model under the synthetic datasets can reach 0.1°. For the natural river datasets, the detection accuracy of the ResNet regression model with residual structure is better than that of VGG16; in terms of the number of model layers, ResNet50-STIV can better balance the detection accuracy and execution efficiency. The detection accuracy in normal scenarios can reach 0.7°, while in flare, turbulence, and rainfall scenarios, it can be controlled within 1.3°, and ResNet50-STIV is better than the existing STIV method. Meanwhile, the maximum relative error is less than that of the instrument method in the surface velocity ratio measurement of multiple scenarios by 12%.