基于MLC-YOLO的折纸动作关键目标检测方法OA

To address the inter-rater and intra-rater variability,as well as the time-consuming nature of manual scoring for the"three-step command"in the Mini-Mental State Examination(MMSE),an object detection method based on MLC-YOLO is proposed to achieve automatic and precise detection of this action.First,multi-level classification object detection is performed on key human body parts and paper,and a same-class maximum confidence selection strategy is adopted to simplify post-processing and ensure the uniqueness of targets of the same class.Second,an adaptive spatial-channel decoupling module is designed to achieve efficient downsampling.Then,Ghost convolution and wavelet transform are introduced into the C3 module to enhance the efficiency and capability of multi-scale feature extraction.Finally,for the small object detection layer,spatial and channel collaborative attention is introduced to improve precision and recall in complex scenes.Experimental results indicate that the mAP95 of the proposed method reaches 61.8%,an increase of 5.6%compared to the original model,while the parameter count is reduced by 51.52%.It proves the effectiveness of the method,offering a new approach for object detection problems involving unique targets and multi-level classification,and providing an effective object detection method for the automatic scoring of the"three-step command"in MMSE.