Abstract:

The traditional methods of segmenting apple leaf images and measuring leaf geometric parameters are moderately accurate but inefficient. To address this problem, an apple leaf image segmentation algorithm based on a deep learning semantic segmentation model and transfer learning was proposed to accomplish efficient and accurate segmentation of apple leaves. The proposed method used LinkNet as the base structure, with the following improvements: ResNet18 was utilized as the backbone network of the encoder and incorporates transfer learning ideas to accelerate model fitting; The number of encoder and decoder blocks was reduced to decrease network complexity; The channel reduction scheme was modified to decrease the parameter quantity in up-sampling; The sub-pixel convolution was introduced to replace the final block to reduce computational costs. Combined with the focal loss, the effectiveness of the improved LinkNet was verified on the standard apple leaf dataset. The experimental results showed that the proposed method achieved a segmentation accuracy of 97.27% and an inference time of 7.82 ms, inference time was decreased by 39.89% compared to the original LinkNet with a slight difference in precision, and the parameter quantity and floating point of operations were significantly reduced. In addition, the inference speed of the improved LinkNet was much faster than that of popular methods such as FCN, U-Net and DeepLabV3+. Therefore, the proposed method could segment the leaf body quickly while better maintaining detailed features such as blade edge serrations. It enabled the efficient and accurate segmentation of apple leaves and provided a novel approach to thinking for fast measurement of leaf area and other geometric parameters.

Key words: deep learning, semantic segmentation, apple leaf, LinkNet, sub-pixel convolution