Oracle bone alphabet detection method for guiding texture feature autonomous learning through LM filter bank

Abstract

The invention discloses an oracle bone alphabet detection method for guiding texture feature autonomous learning through an LM filter bank, and the method comprises the steps: firstly obtaining a rough denoising result of an input image I through employing an Otsu method and a multi-condition connected region filling algorithm; secondly, on the basis of the VGG16 network, introducing a group of trainable convolutional layers at each of the head end and the tail end, and realizing knowledge migration of shallow-layer features and high-layer features through a layer-by-layer freezing training mode; then, introducing a group of Inception sub-networks at the front part of the network, and introducing a group of Leung-Malik direction filter banks at the rear part of the network so as to adapt to the scale and angle changes of characters, and guiding a trainable convolutional layer to effectively obtain difference texture features of a character region; and finally, calculating a score of the region of interest by using a text attention mechanism and the region suggestion sub-network, and determining a text region through the feature dimension reduction sub-network and the region classification sub-network.

Description

technical field [0001] The invention relates to the intersection field of digital image processing and ancient text information processing, especially a kind of device that can effectively resist the interference of dot noises, drilling holes, and burning cracks existing in tortoise shells and animal bones, and can adapt to the interference caused by the engraving process. The oracle bone character detection method is characterized by inconspicuous distribution of rows and columns, complex and changeable line element direction, high accuracy, good robustness, and strong angle adaptive ability. The LM filter group guides the independent learning of texture features. Background technique [0002] As one of the important basic research fields of computational oracle bone science, the fundamental purpose of oracle bone inscription detection is to automatically locate the area position of oracle bone inscriptions on the oracle bone rubbing image with the help of computer vision te...

AI Technical Summary

Problems solved by technology

However, for images to be detected with multiple directions and complex scenes, this type of method cannot obtain effective detection results.

[0006] (2) In terms of regression-based deep learning text detection, Liao Minghui et al. replaced the encoder used to achieve multiple feature fusion in the SSD network structure with a rotation-sensitive regression encoder, and replaced the traditional convolution structure with directional response convolution. The improved deep network structure can extract rotation-sensitive convolution features; in order to overcome the lack of robustness of the SSD network for small target detection, Shi et al. proposed a text detection framework TextBoxes, by adjusting the default aspect ratio of the candidate box, filter scale, which improves the detection performance of long text in the horizontal direction of the deep network; in order to solve the problem that TextBoxes is not ideal for non-horizontal text detection, Liao Minghui et al. replaced the traditional horizontal candidate box with an arbitrary quadrilateral candidate with directional information box, incorporating a regression loss with directional information, thus proposing a TextBoxes++ method for text area detection

Its shortcoming lies in its poor robustness to noise interference; Huang Yongjie et al. proposed an automatic target positioning method for oracle bone rubbing images, which uses target shape estimation as the constraint of the sparse active contour model, and only targets some pixels in the image. Then use the common delineation algorithm to scan out the area with the highest matching degree with the model in the image to be located; He Ying et al. proposed an oracle bone inscription image segmentation algorithm using FCM combined with wavelet transform, using binary wavelet edge detection and The FCM clustering algorithm tracks and fits the edges of the text, and then introduces edge information into the calculation process of the FCM cluster membership to obtain more refined segmentation results.

Its shortcoming is that the detection performance of small or incomplete text is poor; Xing Jici proposed two oracle bone detection methods based on the improved YOLOv3 network architecture, through the integration of simulation noise, anchor frame clustering and other optimization strategies , improving the accuracy of oracle bone script detection

Therefore, the detection technology for modern characters is not suitable for automatic detection of oracle bone inscriptions, and it is impossible to accurately locate oracle bone inscriptions in complex backgrounds

[0014] Second, the existing detection technology for ancient characters can only locate the position of the rubbings in the oracle bone inscription image, but cannot accurately extract the oracle bone characters, and the robustness is poor. Less resistant to disturbances such as burn cracks

However, neither the detection technology for modern characters nor the detection technology for ancient characters can effectively deal with oracle bone inscriptions with complex line element directions

[0016] To sum up, there is currently no one that can effectively resist the interference of dot noises, drilling holes, and burning cracks existing in tortoise shells and animal bones, especially to adapt to the uneven distribution of characters caused by the engraving process. An automatic detection method for oracle bone inscriptions with obvious features, complex and changeable line element directions, high accuracy, good robustness, and strong angle adaptive ability

Images