WiMi Hologram Cloud Inc, a leading global Hologram Augmented Reality (“AR”) Technology provider, announced that its R&D team proposed an asymmetric spectral network algorithm. The algorithm employs asymmetric coordinate spectral spatial feature fusion to provide a novel, end-to-end feature learning method for hyperspectral image classification tasks. The algorithm’s adaptive feature fusion method is capable of extracting discriminative spectral spatial features, and unlike common feature fusion methods, the algorithm is more adaptable to multi-hop connectivity tasks while eliminating the need for manual parameterization.
WiMi’s asymmetric spectral network algorithm solves the spectral noise problem through adaptive feature fusion. The algorithm allows the network to adaptively fuse multiple pieces of information to extract discriminative spectral-spatial features. Unlike traditional feature fusion, this algorithm does not require manual parameterization and is adapted to multi-hop connectivity tasks. This adaptivity helps to efficiently handle complex spectral data and improves the algorithm’s ability to recognize real signals.
In terms of the band correlation problem, the asymmetric spectral network algorithm introduces a coordinate and strip pooling module. Coordinates are used to obtain accurate coordinate and channel information, which helps the network to better understand the spatial structure of the data. Meanwhile, the strip pooling module is used to avoid introducing irrelevant information. The combination of these two techniques makes the network more adaptive and better able to handle the complex band correlations present in hyperspectral images.
WiMi’s asymmetric spectral network algorithm focuses on simplicity, which is to reduce the model complexity with less training time. The algorithm successfully reduces the complexity of the algorithm through an asymmetric learning model and adaptive feature fusion while maintaining high classification performance. This makes the algorithm more suitable for practical application scenarios and provides higher efficiency for hyperspectral image classification tasks.
WiMi’s asymmetric spectral network algorithm focuses not only on static scenes but also on dynamic scenes. Its end-to-end feature learning approach and adaptive feature fusion method enable the algorithm to better adapt to the ever-changing information in hyperspectral images, thus improving the classification accuracy in dynamic scenes. It effectively overcomes the technical challenges in hyperspectral image classification and brings a more efficient and accurate solution.
In addition, it introduces the key technology of asymmetric coordinate spectral spatial feature fusion. The algorithm learns the feature representation of hyperspectral images end-to-end through an asymmetric learning model. Compared to traditional methods, this asymmetric learning approach better captures the complex relationships between pixels, enabling the model to more accurately understand the non-uniformity of the spatial distribution, thus improving the classification accuracy.
SOURCE: GlobeNewswire