A novel approach for apple leaf disease image segmentation in complex scenes based on two-stage DeepLabv3+with adaptive loss

In complex environments, overlapping leaves and uneven light can make pixels of leaf edges difficult to identify, resulting in a poor segmentation performance of the target leaf. In addition, the pixel ratio imbalance between the background area and the target area is the main reason that undermines the accuracy of spot extraction. To address these problems, a novel two-stage DeepLabv3+ with adaptive loss is proposed for the segmentation of apple leaf disease images in complex scenes. The proposed adaptive loss adds a modulation factor to the cross-entropy (CE) loss that can reduce the weight of losses generated by easily classified pixels. Therefore, it allows the model to focus more on hard-to-classify pixels during learning, thus improving segmentation accuracy. The novel two-stage model, consisting of Leaf-DeepLabv3+ and Disease-DeepLabv3+, is named LD-DeepLabv3+. In the first stage of the proposed model, Leaf-DeepLabv3+ is employed to extract the leaves from the complex environment. At this stage, the receptive field block (RFB) and the reverse attention (RA) module are introduced to improve the perception ability of the model for different sizes of blades and their edges. Then, the Disease-DeepLabv3+ is designed to segment disease spots from the erased background leaf images in the second stage of the proposed model. In the Disease-DeepLabv3+, the rates of the dilated convolution in atrous spatial pyramid pooling (ASPP) are adjusted to make it more suitable for extracting smaller targets, and the channel attention block (CAB) is introduced to highlight significant spot information and suppress unimportant information. The experimental results show that the proposed method, which combines LD-DeepLabv3+ with the adaptive loss, reaches 98.70% intersection over union (IoU) for leaf segmentation and 86.56% IoU for spot extraction. Compared with the two-stage model DUNet, the proposed method improves the segmentation accuracy of leaves and spots by 0.93% and 4.27%, respectively. Moreover, the total number of parameters and floating points of operations of the proposed method are only 16.96% and 18.25% of those of DUNet, respectively. Hence, the proposed method can provide an effective solution to extract leaves and disease spots in complex environments and has lower computational costs. This makes it suitable for deployment on mobile devices for applications in agriculture.

Automated summary

To address the difficulties in leaf edge identification and the imbalance of pixel ratios between the background area and the target area, a novel two-stage DeepLabv3+ model with adaptive loss is proposed for apple leaf disease image segmentation. The adaptive loss reduces the weight of losses generated by easily classified pixels, allowing the model to focus more on hard-to-classify pixels and improve segmentation accuracy. The experimental results show that the proposed LD-DeepLabv3+ model with adaptive loss achieves high IoU scores of 98.70% for leaf segmentation and 86.56% for spot extraction. It also outperforms the two-stage model DUNet in terms of segmentation accuracy and computational costs.