Field Weed Recognition Based on an Improved VGG With Inception Module

Field Weed Recognition Based on an Improved VGG With Inception Module

Lifang Fu (Northeast Agricultural University, China), Xingchen Lv (Northeast Agricultural University, China), Qiufeng Wu (Northeast Agricultural University, China) and Chengyan Pei (Northeast Agricultural University, China)
DOI: 10.4018/IJAEIS.2020040101

Abstract

The precision spraying of herbicides can significantly reduce herbicide use, and recognizing different field weeds is an important part of it. In order to enhance the efficiency and accuracy of field weed recognition, this article proposed a field weed recognition algorithm based on VGG model called VGG Inception (VGGI). In this article, three optimizations were made. First, the reduced number of convolution layers to reduce parameters of network. Then, the Inception structure was added, which can maintain the main features, and have better classification accuracy. Finally, data augmentation and transfer learning methods were used to prevent the problem of over-fitting, and further enhance the field weed recognition effect. The Kaggle Images dataset was used in the experiment. This work achieved greater than 98% precision in the detection of field weeds. In actual field, the accuracy could reach 80%. It indicated that the VGGI model has an outstanding identification performance for seedling, and has significant potential for actual field weed recognition.
Article Preview
Top

1. Introduction

Recognizing and removing weed is an important part of field management. Weeds secretions hinder the growth of seedlings and cause the yield reduction in the field (Su, 2017). In addition, weeds are main medium to spread diseases and insect pests (Zhao & Liu, 2019). Weeds compete for environmental resources such as nutrients, sunlight, and water (Hu, 2007). Manual weeding (Xie et al., 2018) and biological weed control (Shabbir et al., 2018) are primitive and traditional methods for weeds control in China (Lv, Dong, Sun, & Li, 2018). On the one hand, the cost for the research work and development of biological weed control is very high (Chen & Qiang, 2015), on the other hand, chemical weeding (Wen, Ying, & Libai, 2007) pollutes the environment (Liu, Wang, & Guan, 2005). There is an urgent need to improve the efficiency of weeds recognition and removing.

The versatility of computer vision has made it a fit tool used in many fields, including precision agriculture (Andrea, Daniel, & Misael, 2017). Machine vision technology was first incorporated into crop detection algorithms to identify crop rows and to segment plants and weeds (Brivot & Marchant, 1996). Researchers developed an automated crop spraying system, and analysised the green histogram of plants images and performed crop and weed segmentation with unspecified classifiers (Aitkenhead et al., 2003). Cao, Wang, Mao, and Zhang (2007) developed a method to identify weeds in field by using position and texture features of drill crops, the method could achieve 93% accuracy. Qiao, He, Zhao, and Tang (2013) developed and tested a method of weed recognition based on multi-spectral images and SVM classifier, this method spent less time than traditional SVM classifier. However, this method had a limitation that the accuracy needs to be improved. Wang and Li (2016) proposed an algorithm to extract the height feature of the image and then fuse it with the texture to identify and obtain a higher recognition rate, the algorithm had an effectiveness of 98.33%. However, the plant detection algorithm presented a limitation that the method pent extremely large computing time. Yan (2018) proposed a method to identify weeds by removing shape and color of images, the goal of which was to prepare for intelligent inter-tillage in fields. In order to improve accuracy of crop and weed identification, a fast field weed identification method was proposed based on the deep convolution network and binary code, the method could achieve 92.7% accuracy (Jiang et al., 2018). However, the effectiveness of the approach was not guaranteed in some production crop scenarios.

Complete Article List

Search this Journal:
Reset
Open Access Articles
Volume 12: 4 Issues (2021): Forthcoming, Available for Pre-Order
Volume 11: 4 Issues (2020)
Volume 10: 4 Issues (2019)
Volume 9: 4 Issues (2018)
Volume 8: 4 Issues (2017)
Volume 7: 4 Issues (2016)
Volume 6: 4 Issues (2015)
Volume 5: 4 Issues (2014)
Volume 4: 4 Issues (2013)
Volume 3: 2 Issues (2012)
Volume 2: 2 Issues (2011)
Volume 1: 2 Issues (2010)
View Complete Journal Contents Listing