Deep Q-Network (DQN) has demonstrated significant application value in path planning. However, several challenges remain in UAV path planning tasks, such as difficulty in perceiving dependencies among the components of the input state, imbalanced experience sampling, and the lack of constraints on action selection. To address these issues, this paper proposes a DQN framework based on a semantic-group self-attention mechanism (SGSA-DQN). First, a semantic-group self-attention Q-network (SGSAQ-network) is designed to improve the accuracy of the network’s Q-value output by capturing the dependencies among semantic component groups in the state vector. Second, a guidance- and perception-based action selection algorithm (GPAS) is proposed to reduce unnecessary exploration in the early training stage and accelerate model convergence. Then, a coverage-constrained experience replay mechanism (CCER) is proposed to balance experience sampling probabilities across different state spaces and enhance model robustness. Finally, a Lyapunov safety constraint (LSC) is introduced to mask unsafe actions and improve the task success rate. Experimental results demonstrate that the proposed method achieves an average success rate of 79.17% in UAV path planning tasks.