A fine segmentation model of flue-cured tobacco’s main veins based on multi-level-scale features of hybrid fusion

Abstract

Flue-cured tobacco (FCT) can be classified into upper (B), middle (C), and lower (X) parts based on characteristics such as the FCT's main veins, leaf shape, color, and thickness. Accurately measuring the geometric parameters of the main veins is crucial for identifying the different parts. However, this task has proven to be challenging. Therefore, segmenting the main veins is a prerequisite to reducing calculation errors and improving the precision of part identification. To obtain enough semantic information and improve segmentation accuracy, we propose a fine segmentation model (MSHF-Net) of FCT's main veins based on multi-level-scale features of hybrid fusion. Firstly, MobileNetV2 with a dilated convolution layer (DMobileNetV2) is selected as the backbone network for feature extraction, which optimizes training and inference speed to minimize computing costs. Subsequently, Hybrid Fusion Atrous Spatial Pyramid Pooling (HFASPP) is designed to be the strengthened backbone module for capturing more high-level semantic information, effectively preventing intermittent segmentation of some main veins. Additionally, considering the low proportion of main vein targets in the original image, the double shallow feature branches (DSFBS) are included to obtain more low-level semantic information. Finally, a channel attention mechanism (ECANet) is added to enhance useful information and eliminate redundant information after the hybrid fusion of high-low-level semantic information, preventing mis-segmentation of regions. Experimental validation demonstrates the efficiency of the MSHF-Net, with parameters of only 7.92 M, thus ensuring minimal computational requirements. The model achieves an impressive mean intersection over union (MIoU) of 85.57% and mean pixel accuracy (mPA) of 93.10% on a diverse test set of FCT parts. When applied to segment main veins in a 2296 × 1548 × 3 tobacco image, the model takes just over 0.1 s. It is noteworthy that none of the 291 randomly segmented tobacco leaf main veins show mis-segmentation, highlighting the model's robustness and practical applicability in various scenarios. These results emphasize the superior segmentation performance of the proposed model, establishing a crucial foundation for accurately discriminating FCT parts.