Monitoring and optimization of flow regimes in aerated stirred tanks is crucially important for process efficiency and product quality. To date, experimentally generated flow maps and correlations are mostly used to classify flow regimes. However, such an approach is highly limited in terms of scalability and transferability. We propose a model for a soft sensor to classify flow regimes of aerated stirred tanks based only on image data. To select an architecture for the model, we compared various modern architectures including LeNet, VGG16, MobileNetV2, Dense121. Of these, LeNet-5 and custom CNN show the best performance. Furthermore, we tested how disturbances of process and light conditions, and the fill level in the reactor affect the classification performance.