基于知识蒸馏的开放世界场景图生成算法OA

Open-world scene graph generation is currently hindered by the intricacy of net-work structures and the excessively large parameter sizes.Consequently,significant com-putational resources are consumed,hindering practical deployment.Therefore,a light-weight open-world scene graph generation(LO-SGG)model was proposed based on a two-stage knowledge distillation framework.Multilevel representations from a teacher model were transferred to a lightweight student model through feature distillation.Thus,the network parameters were reduced,while the deep feature extraction capabilities were pre-served.A multitask collaborative learning paradigm was established.The object detection and relation prediction tasks were jointly optimized.The experimental results demonstrate that LO-SGG maintains open-world generalization capability and achieves a 6%higher re-call than traditional closed-world models.The parameter quantity was reduced to 10%of the complex architectures,with only a 2%decrease in mean precision.The overall per-formance surpasses that of classical benchmark methods by 6%to 19%.