Contour Resampling-Based Garlic Clove Bud Orientation Recognition for High-Speed Precision Seeding

Abstract

Achieving fast and accurate recognition of garlic clove bud orientation is necessary for high-speed garlic seed righting operation and precision sowing. However, disturbances from actual field sowing conditions, such as garlic skin, vibration, and rapid movement of garlic seeds, can affect the accuracy of recognition. Meanwhile, garlic precision planters need to realize a recognition algorithm with low-delay calculation under the condition of limited computing power, which is a challenge for embedded computing platforms. Existing solutions suffer from low recognition rate and high algorithm complexity. Therefore, a high-speed method for recognizing garlic clove bud direction based on deep learning is proposed, which uses an auxiliary device to obtain the garlic clove contours as the basis for bud orientation classification. First, hybrid garlic breeds with the largest variation in shape were selected randomly and used as research materials, and a binary image dataset of garlic seed contours was created through image sampling and various data enhancement methods to ensure the generalization of the model that had been trained on the data. Second, three lightweight deep-learning classifiers, transfer learning based on MobileNetV3, a naive convolutional neural network model, and a contour resampling-based fully connected network, were utilized to realize accurate and high-speed orientation recognition of garlic clove buds. Third, after the optimization of the model’s structure and hyper-parameters, recognition models suitable for different levels of embedded hardware performance were trained and tested on the low-cost embedded platform. The experimental results showed that the MobileNetV3 model based on transfer learning, the naive convolutional neural network model, and the fully connected model achieved accuracy of 98.71, 98.21, and 98.16%, respectively. The recognition speed of the three including auxiliary programs was 19.35, 97.39, and 151.40 FPS, respectively. Theoretically, the processing speed of 151 seeds per second achieves a 1.3 hm2/h planting speed with single-row operation, which outperforms state-of-the-art methods in garlic-clove-bud-orientation recognition and could meet the needs of high-speed precise seeding.

Frequency and Spatial Domain-Based Approaches for Recognition of Indian Sign Language Gestures

“…Contours of an image is resampled (13) to a specified number of points. It is done by removing any redundant dimensions from the contour.…”

Robots and Autonomous Machines for Sustainable Agriculture Production

“…The first category has eleven papers under the following sub-heading: Intelligent sensing for the crop or machine system [1,4,7,9,10,12,16,17,22,23,26]. Currently, a large number of studies focus on deep learning techniques, which have shown their superb impact on robotic sensing applications, as reflected in this issue.…”

“…Currently, a large number of studies focus on deep learning techniques, which have shown their superb impact on robotic sensing applications, as reflected in this issue. Some papers utilized improved YOLO-based [7,23,26], CNN-based [12,16] or RCNN-based [9] methods to developed a detection model for operating target recognition or performance evaluation from the RGB images. To achieve more accuracy, faster and compacter models may be popular due to the cost-effective and feasiblity with low-computing platforms.…”