Convolutional neural network ( CNN) is applied in the image reconstruction of electrical capacitance tomography ( ECT) gradually due to its strong nonlinear fitting ability. Aiming at the hyperparameter regulation problem of CNN model, this paper investigates the effects of the model parameters on the image reconstruction results of ECT. Firstly, a dataset of “ capacitance matrixparticle concentration distribution” with 80 000 random flow patterns and 40 000 typical flow patterns is established with numerical method, then the CNN models with various hyperparameters are trained and validation through the training set in the dataset. The effects of the network hyperparameters, including the network initialization, grid density, number of the convolution kernels, number of the neurons in the fully connected layer and the structure of the hidden layers, on the image reconstruction accuracy are systematically studied. Further, a test dataset composed of 12 000 extra generated flow patterns is utilized to evaluate the performance of the CNN models. Static experiments were performed to compare and analyze the image reconstruction quality with various CNN models. Results demonstrate that the structure of network hidden layers has a relatively great effect on the image reconstruction accuracy, while the network initialization, grid density, number of the convolution kernels, number of the neurons in the fully connected layers have less effect on image reconstruction accuracy.