Containerized edge computing has demonstrated significant potential in handling complex user computation requests, thanks to the lightweight nature and rapid responsiveness of containers. However, under resource-constrained conditions, challenges such as resource waste from redundant deployment of similar containers pose substantial obstacles to efficient resource scheduling. To address these issues, this paper proposes a novel resource scheduling algorithm, Graph Transformer-Enhanced DQN (GTrans-DQN). The algorithm integrates a Graph Transformer into the Q-network, leveraging a message-passing mechanism to capture implicit dependencies between container queues. Additionally, the algorithm employs container reuse techniques to reduce unnecessary resource overhead during container deployment. We also introduce Container Expansion Mapping (CEM) to tackle the mismatch between dynamic changes in the quantity of container resources and the fixed action space of reinforcement learning models. Experimental results demonstrate that our algorithm significantly outperforms baseline in managing complex workflow scheduling tasks, achieving a reduction in server cost of 5.74%-13.60% under large-scale workflows.