Sort:
Open Access Issue
A Pixel–Channel Hybrid Attention Model for Image Processing
Tsinghua Science and Technology 2022, 27 (5): 804-816
Published: 17 March 2022
Downloads:68

In the field of image processing, better results can often be achieved through the deepening of neural network layers involving considerably more parameters. In image classification, improving classification accuracy without introducing too many parameters remains a challenge. As for image conversion, the use of the conversion model of the generative adversarial network often produces semantic artifacts, resulting in images with lower quality. Thus, to address the above problems, a new type of attention module is proposed in this paper for the first time. This proposed approach uses the pixel–channel hybrid attention (PCHA) mechanism, which combines the attention information of the pixel and channel domains. The comparative results of using different attention modules on multiple-image data verify the superiority of the PCHA module in performing classification tasks. For image conversion, we propose a skip structure (S-PCHA model) in the up- and down-sampling processes based on the PCHA model. The proposed model can help the algorithm identify the most distinctive semantic object in a given image, as this structure effectively realizes the intercommunication of encoder and decoder information. Furthermore, the results showed that the attention model could establish a more realistic mapping from the source domain to the target domain in the image conversion algorithm, thus improving the quality of the image generated by the conversion model.

Open Access Issue
Graph Convolutional Network Combined with Semantic Feature Guidance for Deep Clustering
Tsinghua Science and Technology 2022, 27 (5): 855-868
Published: 17 March 2022
Downloads:10

The performances of semisupervised clustering for unlabeled data are often superior to those of unsupervised learning, which indicates that semantic information attached to clusters can significantly improve feature representation capability. In a graph convolutional network (GCN), each node contains information about itself and its neighbors that is beneficial to common and unique features among samples. Combining these findings, we propose a deep clustering method based on GCN and semantic feature guidance (GFDC) in which a deep convolutional network is used as a feature generator, and a GCN with a softmax layer performs clustering assignment. First, the diversity and amount of input information are enhanced to generate highly useful representations for downstream tasks. Subsequently, the topological graph is constructed to express the spatial relationship of features. For a pair of datasets, feature correspondence constraints are used to regularize clustering loss, and clustering outputs are iteratively optimized. Three external evaluation indicators, i.e., clustering accuracy, normalized mutual information, and the adjusted Rand index, and an internal indicator, i.e., the Davidson-Bouldin index (DBI), are employed to evaluate clustering performances. Experimental results on eight public datasets show that the GFDC algorithm is significantly better than the majority of competitive clustering methods, i.e., its clustering accuracy is 20% higher than the best clustering method on the United States Postal Service dataset. The GFDC algorithm also has the highest accuracy on the smaller Amazon and Caltech datasets. Moreover, DBI indicates the dispersion of cluster distribution and compactness within the cluster.

total 2