关键词: UNet, resnet50, 动态加权通道空间注意力, 多尺度边缘检测 迁移学习, 语义分割

Abstract: In order to accurately obtain the minerals under the microscope and provide technical sup⁃ port for the fine segmentation of minerals，a multi-optimized UNet model is proposed to segment the mineral flakes to improve the accuracy of the image segmentation task in the accurate identification direction of min⁃ erals.The improved model encoder integrates the resnet50 structure，and uses the dynamic weighted channel spatial attention and multi-scale edge detection module to maximize the solution to the boundary blur prob⁃ lem of the mineral alteration transition zone.At the same time，the decoder also adds an attention mechanism to make it more accurate to interpret the features obtained by the encoder during decoding，and uses the transfer learning method to improve efficiency.By selecting common rock-forming minerals for model train⁃ ing，the experimental results show that the model can effectively segment minerals from the background ma⁃ trix. The loss rate of mineral segmentation verification value is 18%，the average intersection ratio is 82%， and the accuracy rate is 92%. It can achieve more accurate mineral segmentation and provide technical sup⁃ port for accurate mineral segmentation and recognition.

Key words: UNet, resnet50, dynamic weighted channel spatial attention, multi-scale edge detection, transfer learning, semantic segmentation