The current implementation methods of interior space art design optimization have the problem of imperfect reward function design, which makes the model unable to accurately measure the design quality during the training process, thus affecting the final optimization effect. To solve this problem, this paper adopts a reward reconstruction mechanism that integrates multimodal perception and human feedback based on the public dataset of interior design cases. By introducing a spatial perception module based on image semantic segmentation and a design intent extraction model based on text annotations, a dynamic and learnable composite reward function is constructed. During the training process, the deep reinforcement learning subject continuously optimizes the design scheme through multiple rounds of interaction with the environment, and dynamically adjusts the strategy to match the high-quality design standards by referring to the image semantic understanding results and the design goals extracted from natural language in each round of iteration, so as to achieve a spatial layout optimization that is more in line with human cognition and aesthetics. The experimental results on the Structured3D dataset containing more than 10,000 annotated scenes show that compared with the traditional genetic algorithm, the multimodal perception and reward reconstruction method in this paper improves the passage space retention rate by 17.7%; compared with the unimodal DRL (Deep Reinforcement Learning) method, the basic functional component integrity is improved by 15.2%, and the style classification consistency is improved by 26.2%. Expert blind evaluation confirms that the method proposed in this paper has achieved a comprehensive breakthrough in the rationality of spatial layout, functional integrity and aesthetic consistency.