Scene image visual layout based on deep encoder-decoder network and visual image attention model.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Computational Geometry

Background:

• Generating complex scene layouts is challenged by cross-modal semantic alignment bias and inefficient modeling of dynamic spatiotemporal relationships.
• Existing methods lack sufficient physical constraint awareness and real-time performance, limiting applications in virtual reality and smart cities.

Purpose of the Study:

• To address cross-modal semantic alignment bias, low efficiency in dynamic spatiotemporal relationship modeling, and insufficient physical constraint awareness in scene layout generation.
• To propose a novel framework that enhances physical constraint perception and computational efficiency for complex scene layout generation.

Main Methods:

• A framework utilizing deep encoding-decoding networks and visual graph attention.
• Introduction of overall nested edge detection for optimized multi-scale feature fusion.
• Design of an explicit edge feature modeling strategy to enhance physical constraint awareness.

Main Results:

• Achieved an Intersection-over-Union (IoU) ratio of 0.82 and 94.6% key feature coverage on COCO-Stuff dataset.
• Demonstrated a 38% improvement in training efficiency, 19.5% reduction in memory usage, and controlled energy consumption at 54.9Wh.
• Real-world testing showed IoU difference < 0.03, near professional design capabilities (average 9.39/10), and real-time inference speed of 26.94ms.

Conclusions:

• The proposed framework significantly optimizes cross-modal semantic consistency, multi-scale dynamic interaction modeling, and spatial constraint perception.
• The model provides high-precision and efficient scene layout generation, suitable for virtual reality scene construction and smart city digital design.
• The research validates comprehensive advantages in cross-modal alignment, physical constraint perception, and computational efficiency.