基于深度神经网络的试卷批改标记检测与识别研究OA
Detection and Recognition of Examination Paper Marking Traces Based on Deep Neural Networks
针对试卷批改标记形态多样、尺度差异大和易受印刷文字等复杂背景干扰等问题,提出一种融合频域增强与 Ghost 轻量化的尺度形状解耦试卷批改标记检测算法 FGSD-YOLO.首先,提出频率增强混合注意力模块(FHAT),通过对特征图振幅谱施加通道级自适应加权,增强与批改标记相关的特征响应,抑制印刷文字等背景干扰.其次,引入 GhostConv 轻量化卷积,通过半数通道标准卷积与轻量逐通道线性变换相结合,有效压缩模型参数量与计算开销.最后,提出高斯尺度形状解耦 IoU 损失函数(GSSD-IoU),以高斯衰减形式重构中心距离惩罚并将形状惩罚解耦为宽度与高度 2 个独立约束项,提升小目标与异形标记的回归精度.实验结果表明:FGSD-YOLO 在 Examination Paper Mark Trace 数据集上的 mAP@0.5达到 97.1%,较基准模型提升 4.3 百分点;在 Examination Paper Multiple Choice 数据集上的 mAP@0.5 达到 98.4%,较基准模型提升 3.3 百分点;同时参数量仅为 2.8×106,GFLOPs 为 7.4,实现了检测精度与模型大小的平衡,并具有良好的泛化能力、稳定性和复杂场景适应能力.
To address the challenges posed by diverse morphology,large scale variations,and complex background interference(e.g.,printed text)in examination paper marking traces,FGSD-YOLO,a scale-shape decoupled detection algorithm that integrates frequency-domain enhancement and Ghost-based lightweight design,is proposed in this study.First,a Frequency-enhanced Hybrid Attention module(FHAT)is introduced,which applies channel-wise adaptive weighting to the amplitude spectrum of fea-ture maps,thereby enhancing feature responses relevant to marking traces while suppressing background interference such as printed text.Second,GhostConv lightweight convolution is adopted,combining stand-ard convolution on half of the channels with lightweight pointwise linear transformations,effectively reduc-ing model parameters and computational overhead.Finally,a Gaussian Scale-Shape Decoupled IoU loss(GSSD-IoU)is proposed,which reconstructs the center-distance penalty via Gaussian decay and decou-ples the shape penalty into two independent constraints for width and height,thereby improving regression accuracy for small targets and irregularly shaped marks.Experimental results demonstrate that FGSD-YOLO achieves an mAP@0.5 of 97.1%on the Examination Paper Mark Trace dataset,outperforming the baseline model by 4.3 percentage points,and an mAP@0.5 of 98.4%on the Examination Paper Multiple Choice dataset,exceeding the baseline by 3.3 percentage points.Moreover,the model has only 2.8×106 pa-rameters and the computational cost is 7.4 GFLOPs,achieving a favorable trade-off between detection ac-curacy and model compactness,while exhibiting strong generalization capability,robustness,and adaptabili-ty to complex scenarios.
孙丽丽;冯丽佳
河北省教育考试院,河北 石家庄 050091河北工程技术学院,河北 石家庄 050091
信息技术与安全科学
试卷批改标记频域注意力轻量化网络损失函数优化
examination paper marking tracesfrequency-domain attentionlightweight networkloss function optimization
《机械与电子》 2026 (9)
34-43,56,11
河北省教育科学"十四五"规划2025年度教育决策咨询专项课题(2504039)
评论