A standard convolutional neural network analyzes images using regular sampling grids, which works well for many computer-vision tasks. However, natural medicinal materials often have irregular, curved, and variable shapes that don’t fit neatly into such grids. Fixed convolutions may thus capture unnecessary background along with the relevant material. This patent improves the network’s focus by incorporating deformable convolution modules that learn offsets, enabling the receptive fields to better conform to the shapes in the data. The system also generates multi-scale feature maps and integrates selected features before detection. A coordinate-attention mechanism adds positional context, and deconvolution during feature fusion helps prevent boundary diffusion when multiple materials appear in an image. These innovations aim to enhance the efficiency and accuracy of recognizing medicinal materials, which can look similar despite being different substances. The core concept emphasizes that AI can adapt its feature extraction methods to the irregular structures of objects, rather than forcing all objects into a fixed geometric pattern. Instead of expecting medicinal materials to conform to a rigid neural network shape, this approach allows the sampling pattern to follow the actual evidence and structure naturally present in the data.
Letting AI follow the shape of a medicinal material
Patent number: CN 114818874 A
Inventor(s): Z. Jiang, Z. Cai, L. Bai, Y. Zhang, J. Liu, P. Lü, B. Ye, T. Lan, and D. Zhang
Citation: Z. Jiang, Z. Cai, L. Bai, Y. Zhang, J. Liu, P. Lü, B. Ye, T. Lan, and D. Zhang, “一种识别药材的方法和装置 [Method and apparatus for identifying medicinal materials],” China Patent Application CN 114818874 A, Jul. 29, 2022.
