The paper is devoted to the inverse seismic problem solution. Specifically, the joint problem of the fractured geological region localization and its physical parameters estimation is considered. Mathematically, they are equivalent to the binary classification and scalar regression problems. The modern machine-learning techniques are applied. To generate enough amount of synthetic data, the direct simulation approach was utilized. The seismic wave propagation in the heterogeneous multilayered geological model with the fractured inclusion is considered. To describe the realistic background model the standard isotropic linear elastic model Marmousi-2 is utilized. The continuum model of the layered medium with possible slip planes is applied to treat with the dynamic behavior of the fractured inclusion. These models are formulated in terms of linear and semi-linear hyperbolic partial differential systems of equations. Numerical solutions were obtained by the explicit and explicit-implicit grid-characteristic schemes on the square computational grid. The time step of the explicit algorithm in both cases are limited by the Courant stability condition. To provide the fast computational algorithm, modern multi-core parallel algorithms were developed with OpenMP technology. Initially, the standard single-target convolutional neural network, based on the U-net architecture, was implemented. The reasonable precision of the object localization was achieved. Further, the appropriate multi-target model was developed. This type of models is known for their generalization capability. The training time of both networks are approximately the same. It was clearly demonstrated that this approach significantly improves the solution precision for both optimization problems.