面向生物医学数据库的图像边缘检测与分割技术研究

Research on image edge detection and segmentation technology for biomedical database

针对传统自然图像边缘检测分割算法处理医学图像时存在有精度低、效率差的问题,文中基于Transformer和CNN提出了一种图像边缘分割算法。该算法用窗口多头注意力机制替代原Trans⁃former结构中的多头注意力机制,增强了模型的训练精度。同时,通过改进CNN模型实现了对全局特征和局部特征的兼顾。上采样部分采用渐进式解码器,并利用逐步聚合模块融合全局特征和局部特征,进一步减少了图像分割带来的精度损失。在实验测试中,所提算法的准确率、召回率以及F1值在主流对比算法中均为最优,证明其综合性能较为理想。

In response to the problems of low accuracy and poor efficiency in traditional natural image edge detection segmentation algorithms for processing medical images,this paper proposes an image edge segmentation algorithm based on Transformer and CNN.This algorithm replaces the multi head attention mechanism in the original Transformer structure with a window multi head attention mechanism,enhancing the training accuracy of the model.Meanwhile,by improving the CNN model,it achieved a balance between global and local features.The upsampling part adopts a progressive decoder and utilizes a progressive aggregation module to fuse global and local features,further reducing the accuracy loss caused by image segmentation.In the experimental testing,the accuracy,recall,and F1 value of the proposed algorithm were all the best among mainstream comparison algorithms,proving that its comprehensive performance is relatively ideal.