Tian Wen-Fei, Chen Ming, Shu Zhong, Tian Xue-Jun
Faculty of Innovation Engineering, Macau University of Science And Technology, Macao, China.
Jingmen Ecological Environment Monitoring Center of Hubei Provincial Department of Ecology and Environment, Jingmen, Hubei, China.
Sci Rep. 2025 Apr 28;15(1):14880. doi: 10.1038/s41598-025-99163-w.
Source images and predicted target images differ in image features. When heterogeneous transfer learning is applied some difficulties and further issues appear. For example, noise in image recognition appears and is required to be reduced. The Image Feature Data Learning and the Definition of Image Feature Data Consistency modules adopt the normalization layer of a neural network to extract 3 types of features, namely, global features, feature space, and feature labels. A noise reduction method, Rudin-Osher-Fatemi, is implemented. Thus, the Image Feature Data Association Fusion Heterogeneous Transfer Learning Model is proposed. Also, a correlation coefficient is computed for image feature vectors, and effective correlation mapping matrices are constructed through multi-dimensional vectorized correlation. Then, the feature vectors and correlation coefficients are aggregated using the Batch Normalization Layer to assess correlations between image features. Furthermore, to check the variance between the features of the source images and the target images to be minimum and the common space of the transfer mapping features to be maximum, the Definition of Image Feature Data Consistency deals with controlling parameter separability by designing the constraint matrix with minimum variance score for the source images and the target images. Finally, the regularization of the transfer mapping matrix is carried out to create the loss function to consistently train the image features to construct the heterogeneous transfer learning module. When the transfer learning weight matrix is attained, the consistent constraint strategy of the image features is introduced to update image features in real time. Besides, the Gaussian kernel function is employed to control the generated noise in transfer learning. The results indicate that the SNR is greater than 35dB and the edges in the image feature map are clearer and contain less noise in the Image Feature Data Learning module with the Rudin-Osher-Fatemi denoising strategy.
源图像和预测的目标图像在图像特征上存在差异。当应用异构迁移学习时，会出现一些困难和进一步的问题。例如，图像识别中会出现噪声，需要进行降噪处理。图像特征数据学习和图像特征数据一致性定义模块采用神经网络的归一化层来提取3种类型的特征，即全局特征、特征空间和特征标签。实现了一种降噪方法，即鲁丁-奥舍-法特米方法。因此，提出了图像特征数据关联融合异构迁移学习模型。此外，计算图像特征向量的相关系数，并通过多维矢量化相关构建有效的相关映射矩阵。然后，使用批归一化层聚合特征向量和相关系数，以评估图像特征之间的相关性。此外，为了检查源图像和目标图像的特征之间的方差最小，以及迁移映射特征的公共空间最大，图像特征数据一致性定义通过为源图像和目标图像设计具有最小方差分数的约束矩阵来处理控制参数可分离性。最后，对迁移映射矩阵进行正则化，以创建损失函数，从而一致地训练图像特征以构建异构迁移学习模块。当获得迁移学习权重矩阵时，引入图像特征的一致约束策略以实时更新图像特征。此外，采用高斯核函数来控制迁移学习中产生的噪声。结果表明，在采用鲁丁-奥舍-法特米去噪策略的图像特征数据学习模块中，信噪比大于35dB，图像特征图中的边缘更清晰，噪声更少。
