Traditional image-based visual servoing (IBVS) algorithms rely heavily on accurate depth estimation of feature points to construct the image Jacobian matrix, which is often impractical in real-world applications. In contrast to conventional Jacobian-based methods, this work introduces a paradigm-shifting approach by leveraging a genetic algorithm-enhanced backpropagation neural network to bypass depth estimation entirely. The proposed method directly learns the nonlinear mapping from feature position errors to UAV motion velocities, eliminating the need for explicit Jacobian matrix computation or depth-dependent approximations. The principal innovations involve developing a data-driven adaptive framework that successfully accomplishes visual servoing control under depth-agnostic conditions, combined with genetically optimized network training to systematically evade local minima. Simulations validate that the new method achieves superior robustness against depth uncertainty compared to classical IBVS, while maintaining competitive positioning accuracy.