Evaluation And Detection Of Calligraphy Copy Based On Deep Learning

Calligraphy copy is the most important part of practicing,but most researches are based on the extraction of manual features and rely on rich prior knowledge.The existing machine evaluation methods lack of unified quantitative standards,slow speed,poor robustness which is not suitable for real-time evaluation.Traditional calligraphy evaluation is slow,poor in robustness and difficult to capture local differences between copy and template characters.By transforming the evaluation into a comparison problem,the scene is limited to template matching and visual results can be given in real time based on the object detection method.Original YOLOX is simplified so that the speed is increased from 59 frames per second to 132 frames per second(an increase of 73 frames per second).The ablation experiment verifies that parallel network structure will reduce the speed of the model and bring negative benefits to the precision in small models.For a large number of small targets in the data sets,the backbone network was improved by integrating the parameter-free self-attention mechanism into the backbone.It helps model strengthen the attention to small objects.The accuracy on test sets is increased from 80.64%to 87.57%(an increase of 6.93%),and only 17 frames per second(132 frames per second to 115 frames per second)are lost.In addition,the algorithm designed by prior rules which can cover possible situations in practicing can synthesize some data sets.The experiments have proved this method effective.For the detection of calligraphy characters,the self-attention model has higher fine granularity which may get more accurate cropped results.By designing detailed ablation experiments,the model gets a suitable training strategy on challigraphy sets.For the size of images,the adaptive mask is proposed to shield the zero filling in the calculation process so that the model is more stable in training processes.Under the same conditions,the performance is improved from 69.31%to 76.65%(an increase of 7.34%).Finally,the accuracy of the model is improved to 86.10%with pretrained model.