Abstract
Powdery mildew is a common crop disease and is one of the main diseases of cucumber in the middle and late stages of growth. Powdery mildew causes the plant leaves to lose their photosynthetic function and reduces crop yield. The segmentation of powdery mildew spot areas on plant leaves is the key to disease detection and severity evaluation. Considering the convenience for identification of powdery mildew in the field environment or for quantitative analysis in the lab, establishing a lightweight model for portable equipment is essential. In this study, the plant-leaf disease-area segmentation model was deliberately designed to make it meet the need for portability, such as deployment in a smartphone or a tablet with a constrained computational performance and memory size. First, we proposed a super-pixel clustering segmentation operation to preprocess the images to reduce the pixel-level computation. Second, in order to enhance the segmentation efficiency by leveraging the a priori knowledge, a Gaussian Mixture Model (GMM) was established to model different kinds of super-pixels in the images, namely the healthy leaf super pixel, the infected leaf super pixel, and the cluttered background. Subsequently, an Expectation–Maximization (EM) algorithm was adopted to optimize the computational efficiency. Third, in order to eliminate the effect of under-segmentation caused by the aforementioned clustering method, pixel-level expansion was used to describe and embody the nature of leaf mildew distribution and therefore improve the segmentation accuracy. Finally, a lightweight powdery-mildew-spot-area-segmentation software was integrated to realize a pixel-level segmentation of powdery mildew spot, and we developed a mobile powdery-mildew-spot-segmentation software that can run in Android devices, providing practitioners with a convenient way to analyze leaf diseases. Experiments show that the model proposed in this paper can easily run on mobile devices, as it occupies only 200 M memory when running. The model takes less than 3 s to run on a smartphone with a Cortex-A9 1.2G processor. Compared to the traditional applications, the proposed method achieves a trade-off among the powdery-mildew-area accuracy estimation, limited instrument resource occupation, and the computational latency, which meets the demand of portable automated phenotyping.
Highlights
We propose a lightweight powdery mildew spot segmentation model based on the super-pixel segmentation method and hybrid Gaussian clustering method
The first column is the original image, the second column is the manually labeled image, the third column is the result obtained by the U-net method, the fourth column is the segmentation result of the K-mean clustering method, the fifth column is the segmentation result of the maximum interclass variance, and the last column is the segmentation result of the method proposed in this paper
By analyzing the images of these three numbers, it was found that the diseased areas of these samples were smaller, and the clustering effect of the Gaussian Mixture Model became poor, resulting in some pixel points of healthy leaf areas being identified as powdery-mildew diseased areas, leading to lower recognition accuracy
Summary
During the growth of crops, many diseases can directly affect growth. Powdery mildew is one of the common fungal diseases that infects plant leaves and affects their photosynthesis and, yield. Manual identification of the degree of disease [1]. A new method is needed to replace the manual detection of diseases. Current image-based phenotyping methods mainly include chlorophyll-fluorescenceimaging-based methods, hyperspectral-imaging-based methods, thermal-imaging-based methods, and visible-light-image-based methods [2].
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
