摘要
为了提高手术器械语义分割识别算法的分割准确率和分割速率,使其更适应实际应用,本文对DeepLab V3+网络进行优化改进,在ASPP模块将3×3空洞卷积分解成3×1和1×3的空洞卷积,减少冗余参数,在解码端使用深度可分离卷积替代普通卷积,减少信息丢失,保留更多边缘特征。实验结果表明,改进后的网络能实现对手术器械的准确分割,且满足实时性要求,可以应用于手术器械管理系统。
In order to improve the segmentation accuracy and speed of the surgical instrument semantic segmentation recognition algorithm,and make it more suitable for practical application,in this paper,the DeepLab V3+network has been optimized and improved.The 3x3 Atrous convolution has been decomposed into 3x1 and 1x3 Atrous convolution in ASPP module to reduce redundant parameters.Depthwise separable convolution is used to replace ordinary convolution at decoding end to reduce information loss and retain more edge features.The experimental results show that the improved network can achieve accurate segmentation of surgical instruments,and meet the real-time requirements,which can be applied to the surgical instruments management system.
作者
董沛君
陶青川
DONG Peijun;TAO Qingchuan(College of Electronics and Information Engineering,Chengdu 610065)
出处
《现代计算机》
2021年第13期71-75,共5页
Modern Computer