Maize seedling detection under different growth stages and complex field environments based on an improved Faster R–CNN. (August 2019)

• Record Type: Journal Article
• Authors: Quan, Longzhe; Feng, Huaiqu; Lv, Yingjie; Wang, Qi; Zhang, Chuanbin; Liu, Jingguo; Yuan, Zongyang
• Is Part Of: Biosystems engineering. Volume 184(2019)
• Page Start: 1
• Page End: 23
• Publication Date: 2019-08
• Subjects: Maize seedlings -- Faster R–CNN -- Pretrained networks -- Field robot platform -- Machine vision -- Deep learning; Bioengineering -- Periodicals; Agricultural engineering -- Periodicals; Biological systems -- Periodicals; Génie rural -- Périodiques; Systèmes biologiques -- Périodiques; 631
• DOI: 10.1016/j.biosystemseng.2019.05.002
• Languages: English
• ISSNs: 1537-5110

Abstract

Abstract : This paper presents an improved Faster R–CNN model for a field robot platform (FRP) aimed at automatically extracting image features and quickly and accurately detecting maize seedlings during different growth stages under complex field operation environments, with the goal of preparing for intelligent inter-tillage in maize fields. A FRP with five industrial USB cameras for data collection was used to capture a large number of sample images. The shooting angle range of the industrial USB cameras is 0–90°. The photographs were used to create an image database containing twenty thousand images of soil, maize and weeds. Ten selected pretrained networks were used to replace the network of the CNN feature computing component of the classic Faster R–CNN. A Faster R–CNN with VGG19 processed by the pretrained networks method is proposed. The Faster R–CNN algorithm used in this work represents a deep learning architecture that distinguishes maize seedlings and weeds under three field conditions: Full-cycle, Multi-weather and Multi-angle. This work achieved greater than 97.71% precision in the detection of maize seedlings with respect to soil and weeds. The precision rate of six-leaf to seven-leaf maize seedlings was 2.74% lower than that of the total test set. The precision rate under sunny conditions was 1.97% lower than that of the total test set. The precision rate of an angle shot of 0° was 0.95% lower than that of the total test set. The proposed model has significant potential for autonomous weed and maize classification under actual operating conditions. Highlights: 10 pretrained networks replaced the network of CNN feature of the Faster R–CNN. Improved Faster R–CNN network using pretrained networks method proposed. Detection of maize seedlings at different growth stages in the fields is realised. Detection under a variety of weather and camera conditions realised.