用于低功耗图像识别设备的轻量化集成学习方法OA

With technological advancements,power grid safety supervision increasingly relies on AI-based image processing for environmental information analysis.However,as model performance improves,the computational burden inevitably increases significantly.In order to improve the model performance of the existing mainstream networks and apply them in low-power devices,starting from ensemble learning,we propose and implement a lightweight ensemble learning method.Firstly,by optimizing the architecture of individual learner units,we develop an ensemble learning framework compatible with general-purpose image recognition networks.Secondly,we introduce a tailored training strategy aligned with this ensemble learning architecture.This strategy employs a novel error-type-based generalization method for image augmentation,enabling differentiated training across distinct learners.Finally,the implementation on MobileNetV2 achieved a classification accuracy of 95.58%on the TrashNet dataset.Compared to standalone Mo-bileNetV2,the ensemble approach improved performance by 1.16 percentage points while adding only 0.207 GFLOPs in computational overhead and 9.4 M parameters.Comprehensive comparative studies and statistical analysis demonstrate that the proposed method effectively enhances the performance of a general-purpose model while maintaining competitiveness against state-of-the-art approaches.