Multiple kernel-enhanced encoder for effective herbarium image segmentation

Abstract

The neural network proposed here specializes in herbarium image segmentation. The encoder of the proposed model contains multiple kernels of different sizes to address the complex structures of plant components, such as tangled roots and stems. By employing multiple kernel sizes, the convolution block enables multiscale learning, which is underexplored in previous approaches. This design effectively extracts and fuses local and global features, enabling both broad and narrow perspectives on complex structures within herbarium images and thereby improves segmentation performance. The experimental results demonstrate that the proposed model outperforms three conventional models. The source code can be accessed at