Machine Learning Algorithm Based Plant Root Contour Extraction in Edge Devices

The status of plant roots serves as a crucial metric reflecting the plant’s growth state. Extracting and processing the contour of plant roots aids in analyzing the growth status effectively. In this study, high-resolution root images of plants were obtained using embedded optical imaging devices, transferred to edge computing nodes for image processing. By fine-tuning machine learning parameters, an artificial neural network model was developed to integrate Holistically-Nested Edge Detection (HED) and Canny edge detection algorithms.This paper presents a methodology for model training and application based on convolutional neural networks on edge devices, typically single-board computers with limited memory (such as Raspberry Pi), addressing issues with traditional Canny edge detection’s sensitivity to salt-and-pepper noise and susceptibility to producing false contours due to image gradient changes. The algorithm leverages HED, employing machine learning solutions through a large collection of plant root image data for pre-training the model, effectively filtering images while retaining accurate edge information. Furthermore, the output of this method serves as input for the Canny edge extraction algorithm, utilizing an adaptive thresholding algorithm to generate clear plant root contour images.Experimental results demonstrate that the combined approach using embedded artificial neural network models and Canny contour extraction reduces device performance requirements while improving the quality of root contour extraction. This research contributes to a more accurate understanding of morphological features of plant root systems, providing comprehensive imaging data for plant root studies.