Fully Convolutional Network Based Skeletonization for Handwritten Chinese Characters

Abstract

Structural analysis of handwritten characters relies heavily on robust skeletonization of strokes, which has not been solved well by previous thinning methods. This paper presents an effective fully convolutional network (FCN) to extract stroke skeletons for handwritten Chinese characters. We combine the holistically-nested architecture with regressive dense upsampling convolution (rDUC) and recently proposed hybrid dilated convolution (HDC) to generate pixel-level prediction for skeleton extraction. We evaluate our method on character images synthesized from the online handwritten dataset CASIA-OLHWDB and achieve higher accuracy of skeleton pixel detection than traditional thinning algorithms. We also conduct skeleton based character recognition experiments using convolutional neural network (CNN) classifiers on offline/online handwritten datasets, and obtained comparable accuracies with recognition on original character images. This implies the skeletonization loses little shape information.

Query Pixel Guided Stroke Extraction with Model-Based Matching for Offline Handwritten Chinese Characters

“…Recently, FCN based skeletonization has been proven to outperform the above methods remarkably [31] , but for training FCN, it is infeasible to label skeleton pixels for millions of offline handwritten samples [32] . Therefore, our previous work proposed to generate annotated skeleton data from online handwritten samples [17] .…”

“…Fortunately, online handwritten characters record strokes by (x, y ) -coordinate sequences [32] , which can be viewed as the ideal skeletons of synthesized images (generated by dilating the stroke skeletons) [17] . Thus, we generate a large number of synthesized character images from online handwritten characters to train all the models in this paper.…”

“…As shown in Fig. 3 , each online sample records its strokes in writing orders, thus, all stroke images can be generated by dilating the strokes of a plotted online character image with appropriate control of stroke width (randomly sampled) and edge smoothness [17] . This step outputs offline stroke images s 1 , ..., s n , which are generated in writing order and listed in the second row of Fig.…”

“…When processing offline handwritten characters with variable stroke widths and cursive writings, analysis on skeletons is more reliable [16][17][18] . On skeletons, the intersection relationship between strokes can be detected at cross/fork points, and stroke segments can be traced.…”

“…In the proposed method, the components SkeNet and CrossNet (which share the backbone but have different output branches) have been reported in our previous conference paper [2,17] . The main contribution of this paper is to present the comprehensive scheme of handwritten stroke extraction, as well as the components of query pixel guided stroke extraction by PathNet and model-based stroke matching.…”