An Intelligent Marking Algorithm Based On Lightweight Network And Attention Mechanism

In the current education industry,intelligent marking has great advantages in time and cost compared with manual marking,while the intelligent marking algorithm based on cursor reader has many limitations in cost and scenes.Then,this paper designs an end-to-end intelligent marking algorithm based on the deep neural network for arithmetic papers,which can complete the intelligent marking of arithmetic papers in a short time.The main research contents are as follows:(1)In order to reduce model parameters and suppress the inconsistency of feature fusion,a marking detection algorithm based on depth convolution and coordinate attention is proposed.Firstly,the algorithm selects YOLOv5 as the basic framework and uses Mobile Next as the backbone by virtue of deep convolution to improve the inference speed of the model.Secondly,by introducing the coordinate attention mechanism in the early stage of feature fusion,the network can adaptively learn the key feature information of each scale and alleviate the information conflict caused by feature fusion of different scales.Finally,the multi-scale features and NMS technology are used for joint prediction.Experiments show that the algorithm can effectively improve the detection accuracy of the scoring model while reducing parameters.(2)Considering the problem of character inclination and sticking of arithmetic test questions,a marking recognition algorithm based on spatial transformation and enhanced attention is proposed.The algorithm selects CRNN as the basic framework of marking recognition.By introducing the STN at the beginning of the network,the position of the test image is corrected by the affine transformation to accelerate the convergence of the model.Then,an enhanced attention mechanism is added to the feature extraction network to make the network focus on the adhered character area.Finally,the CTC decoding mechanism is used to obtain the semantic information.The results show that the algorithm significantly improves the recognition accuracy of the marking model at the cost of introducing a small amount of parameters.(3)Due to the limitations of model deployment,a lightweight scoring algorithm applied to the mobile device is proposed.The algorithm firstly selects the ESNet which focuses on the mobile detection as the backbone to improve the inference speed of the model and replaces the CSP module in the feature fusion network into a more lightweight DWBlock.Then the multi-scale prediction is reduced from three scales to two scales.Finally,the network is compressed by lowering the depth and width multiplier for the deployment of the mobile device using the ncnn framework.After testing,the algorithm realizes the real-time inferencing of the scoring model on the mobile device at the cost of losing a bit of precision.