基于改进卷积神经网络的图像数字识别方法研究

Research on Image Digital Recognition Method Based on ImprovedConvolutional Neural Network

针对试卷分数的统计问题,采用一种带有特殊分值框的试卷,并提出了一种基于改进卷积神经网络的识别统计方法。首先基于YOLO目标检测算法对分值框进行定位,并引入膨胀卷积模块丰富感受野、调整边框损失函数、提高收敛速度,然后基于ResNet卷积神经网络对分数进行识别,并融合注意力机制提高特征提取能力。实验结果表明,经改进的模型对1 000份试卷中题目分数的识别准确率为99.2%,可以准确、高效地识别试卷图像中的分数。

Aiming at the statistical problem of test paper scores,this paper proposes a recognition statistics method based on improved convolutional neural network by using a test paper with a special score box.Firstly,the YOLO object detection algorithm is used to locate the score boxes,and the dilated convolutional module is introduced to enrich the receptive field and adjust the border loss function of the frame to improve the convergence rate.Then,the score is recognized based on ResNet convolutional neural network,and the attention mechanism is integrated to improve the feature extraction ability.The experimental results show that the improved model has a recognition accuracy of 99.2%for question scores in 1000 test papers,and can accurately and efficiently recognize scores in test paper images.