Image Segmentation of Plant Leaves in Natural Environments Based on LinkNet

Abstract

The conventional approach to plant leaf image segmentation for subsequent measurement of leaf geometric parameters, while reasonably accurate, exhibits lower efficiency. To address this challenge, a plant leaf image segmentation algorithm based on deep learning semantic segmentation models and transfer learning is proposed to achieve rapid and precise leaf segmentation. The presented method adopts LinkNet as its foundational network structure and introduces four key enhancements: leveraging ResNet18 as the backbone encoder network to expedite model fitting through transfer learning; reducing the number of encoding and decoding blocks to decrease network complexity; refining channel reduction strategies to minimize parameters in upsampling; and implementing subpixel convolution for upsampling to lower computational load. Integrated with the focal loss function, this improved LinkNet network is applied to a standard leaf dataset. Experimental outcomes indicate that the proposed algorithm achieves a segmentation accuracy of 97.27%, comparable to the original LinkNet; inference time is merely 7.82 ms, a 39.89% reduction compared to the original network; a substantial decrease in model parameters and floating-point operations; and a significantly faster inference speed of the improved network compared to FCN, U-Net, DeepLabV3+, and similar networks. This algorithm not only swiftly segments the main body of leaves but also effectively preserves details such as leaf edge serrations. It truly enables efficient and precise segmentation of plant leaves, offering a novel approach for rapidly measuring leaf area and other geometric parameters.