翻译：1 检测难点文献甲骨文检测与普通的文字检测不同具有以下难点：1文献甲骨文检测属于小目标检测所以每个需检测的甲骨文图像包含的信息过少因而它的判别性特征过少。同时小目标检测需要对检测框位置坐标的预测非常精准因为小目标检测本身目标包含像素数少与普通目标检测相比当预测坐标发生相同的偏移时小目标检测的IOU误差更大。2文献图像中文字负样本数量较多而甲骨文字正样本数量较少。这导致网络不能很好的学习到正样

Challenges in detection The detection of oracle bone inscriptions in literature differs from ordinary text detection and presents the following challenges: (1) Detection of oracle bone inscriptions belongs to small object detection. Therefore, each image of oracle bone inscription to be detected contains too little information, resulting in a lack of discriminative features. At the same time, small object detection requires very accurate prediction of the detection box position coordinates, as small object detection itself contains fewer pixels than ordinary object detection. Compared with ordinary object detection, when the predicted coordinates undergo the same offset, the IOU error of small object detection is greater. (2) There are more negative samples (characters in literature images that are not oracle bone inscriptions) than positive samples (oracle bone inscriptions) in the literature images. This results in the network not being able to learn the features of positive samples well. (3) The original image resolution is too high. Through pre-experimental results, it was found that if the image clarity is ensured, directly putting literature images with a resolution of about 4000 to 6000 pixels into deep learning network training will result in training loss of NAN, failure to detect target information, or memory overflow, as shown in Figure 5