Due to the large volumes of seismic in the industry, there is a constant effort to develop automatic or semi-automatic tools for picking horizons, faults etc. The variety of convolution neural networks proposed for automatic interpretation of seismic images, especially for faults detection. In this paper, we test different CNN models for faults detection and derive the key neural network parameters that influence on the faults localization. We aim to derive the CNN parameters, that allows to detect thin area of the fault and balanced detection of the unmarked faults. We provide the experiments on the open F3 Northen Block dataset, which is popular for benchmarking of the machine learning solutions in seismic interpretation. The best of the tested models allows to highlight the unmarked faults. The accuracy of this model for test and validation dataset is 0.97/0.96, precision, recall and f1 score for faults and background classes are 0.55/0.87, 1.00/0.98, 0.68/0.99, the Jaccard similarity score is 0.94.