特定課題制度(学内資金)
特定課題制度(学内資金)
-
2023年
概要を見る
During this research period, we investigated developing a deep learningmodel for hyperspectral image (HSI) classification tasks based on the popularTransformer architecture. By identifying shortcomings in the existingTransformer model, we proposed a new perspective for integrating RecurrentNeural Networks (RNNs) with Transformers for complementarity. The proposalconsisted of three components: 1) RNN-Transformer encoder, 2) Soft MaskedSpectral-Spatial-Based Self-Attention (SMSA), and 3) Multiscanning FusionTransformer. In this case, the transaction paper titled “Multiscanning-based RNN-Transformer for HyperspectralImage Classification” was accepted by the IEEETransactions on Geoscience and Remote Sensing (TGRS). Compared withbaseline methods, this work achieved a 6%~11% accuracy improvement. Moreover,compared with other state-of-the-art methods, our work obtained a 1%~5%accuracy improvement and saved almost 50% of processing time with almost 40%model size reduction.Meanwhile, the idea of the “multiscanning strategy” was extended into another field, image compression, which was studiedby our coworkers. The paper, titled “Learned Image Compression with Multi-Scan Based Channel Fusion,” was accepted by the International Conference onImage Processing (ICIP) in 2023. This work verified the effectiveness ofthe “multiscanning strategy” and showed the general attribute of this idea. We hoped to furtherdevelop this concept into other research fields.Furthermore, to facilitate the multiscanning strategy into a 3D version, we proposed a cubed 3D-multiscanning strategy. The manuscript, titled "Segmented Recurrent Transformer with Cubed 3D Multiscanning Strategy for Hyperspectral Image Classification", is accepted at 26 March, by IEEE Transactions on Geoscience and Remote Sensing (TGRS) .
-
2022年 Sei-ichiro Kamata
概要を見る
I published two papers in international conferences - the 26th International Conference on Pattern Recognition (ICPR) and the International Conference on Image Processing (ICIP) in 2022. The first paper [1], presented a novel approach to designing a unified spectral-spatial Transformer for hyperspectral image classification. Specifically, I proposed a cascaded integration of the spectral vision Transformer with the spatial pyramid vision Transformer, along with a cross-scale fusion module. Moreover, I introduced a local-global encoder in the spatial domain, which validates the effectiveness of incorporating local features into the Transformer model. Overall, my paper contributed to the advancement and practicality of using a pure vision Transformer-based model for hyperspectral image classification. The second paper [2] proposed a new approach for addressing hyperspectral image classification by leveraging the 3D configuration of a vision Transformer, which enabled simultaneous correlation of spectral and spatial features. To this end, I introduced a novel 3D coordinate positional embedding method that distinguished the relative distances among all hyper-cubes resulting from the 3D partition operation. I also designed a local-global feature combination approach that seamlessly integrates with the 3D configuration of the vision Transformer. Furthermore, we presented our research at two conferences and received positive feedback.