Feature representation learning and combinatorial optimization method of visual design elements

Abstract

Aiming at the challenge of automatic combinatorial optimization of visual design elements, this paper proposes a hybrid framework that combines multi-modal feature representation learning and hierarchical combinatorial optimization. By constructing a dual-channel graph neural network (GNN) encoder, the visual content features and spatial relationship features of design elements are extracted synchronously, and the collaborative representation of multimodal information is realized. The combinatorial optimization module based on reinforcement learning (RL) models the design process as a Markov decision process (MDP), designing a composite reward function to balance aesthetic principles, functional constraints, and user intent. A hybrid optimization strategy combining genetic algorithms (GA) and gradient descent is further introduced, utilizing the global exploration capability of GA to quickly locate potential solutions and integrating gradient refinement for local fine-tuning, effectively solving the issue of local optima in high-dimensional search spaces. The experimental results show that this method is significantly superior to the baseline model in the component relationship prediction task, and achieves 9.0% rule violation rate and 0.81 intention coincidence degree in the layout generation task, and is unanimously recognized by professional designers and ordinary users. This research provides a new paradigm for data-driven intelligent design generation.