Error collection method, device and equipment based on deep learning

The invention provides an error collection method, device and equipment based on deep learning. The method comprises: acquiring corrected homework or test paper images containing multiple questions; carrying out topic segmentation on the obtained image by adopting a pre-trained topic coordinate-based recognition model to obtain a coordinate region corresponding to each topic; and identifying the correction trace in the coordinate area corresponding to each question according to a pre-trained detection model based on the correction trace so as to obtain the right and wrong judgment of the question. 本发明提供一种基于深度学习的错题收集方法、装置及设备，所述方法包括：获取已完成批改的且包含有多道题目的作业或试卷图像；采用预先训练得到的基于题目坐标的识别模型对获取的图像进行题目切割，以获得每道题目所对应的坐标区域；根据预先训练得到的基于批改痕迹的检测模型对每道题目所对应的坐标区域内的批改痕迹进行识别，以获得该道题的对错判定。