深度特征融合与重构的微纤维识别算法

Microfiber recognition algorithm based on deep feature fusion and reconstruction

针对传统卷积神经网络分类识别微纤维存在特征判别不明显的问题,构建了一种深度特征融合与重构的网络对其进行分类与识别。将卷积与深度可分离卷积特征进行融合,加强层间信息交流,提高特征判断指向能力,并在上采样之前分配通道和空间的权重进行特征重构,利用通道注意力与空间注意力相结合的策略使网络在学习的过程中将注意力集中在关键的特征信息处,同时,跳跃连接增加原始特征图,缓解拟合现象,强化微纤维区域关键特征信息,提升微纤维图像识别网络模型的表达能力和学习能力,从而改善微纤维识别效果。实验结果表明,微纤维识别率达到98.77%,通过特征图可视化进一步分析了特征融合与重构的作用。所构建的方法准确率高、泛化能力好,为微纤维分类识别提供了一种新的方案。

In view of the fact that the feature discrimination of the traditional convolutional neural network is not obvious in the microfiber classification and recognition, a deep feature fusion and reconstruction network is constructed to perform classification and reorganization. The convolution and depth separable convolution features are fused to strengthen information exchange between layers,so as to improve feature judgment and pointing ability of the network. The channel and space weights are assigned for feature reconstruction before upsampling. The strategy of combining channel attention with spatial attention is used to enable the network to focus the attention on the key feature information in the learning process. At the same time,jump connections is used to increase the original feature map,alleviate the fitting,strengthen the key feature information of the microfiber area,and improve the expressive ability and learning ability of the microfiber image recognition network model,so as to improve the microfiber recognition effect. The experimental results show that the recognition rate of microfibers can reach 98.77%,the function of feature fusion and reconstruction is further analyzed by feature map visualization,and the constructed method has high accuracy and good generalization ability,which provides a new scheme for microfiber classification and recognition.