A coarse-to-fine network for aphid recognition and detection in the field. (November 2019)

• Record Type: Journal Article
• Authors: Li, Rui; Wang, Rujing; Xie, Chengjun; Liu, Liu; Zhang, Jie; Wang, Fangyuan; Liu, Wancai
• Is Part Of: Biosystems engineering. Volume 187(2019)
• Page Start: 39
• Page End: 52
• Publication Date: 2019-11
• Subjects: Aphid Detection -- Aphid Recognition -- Convolutional Neural Network -- Coarse-to-Fine Network; Bioengineering -- Periodicals; Agricultural engineering -- Periodicals; Biological systems -- Periodicals; Génie rural -- Périodiques; Systèmes biologiques -- Périodiques; 631
• DOI: 10.1016/j.biosystemseng.2019.08.013
• Languages: English
• ISSNs: 1537-5110

Abstract

Abstract : In agriculture, aphids are one of the most destructive pests, responsible for major reductions in wheat, corn and rape production leading to significant economic losses. However, manual pest recognition approaches are often time-consuming and laborious for Integrated Pest Management (IPM). In addition, the existing pest detection methods based on Convolutional Neural Network (CNN) are not satisfactory for small aphid recognition and detection in the field because aphids are tiny and often in dense distributions. In this work, a two-stage aphid detector named Coarse-to-Fine Network (CFN) is proposed to address these problems. The key idea of our method is to develop a Coarse Convolutional Neural Network (CCNN) for aphid clique searching as well as a Fine Convolutional Neural Network (FCNN) for refining the regions of aphids in the clique. Specifically, The CCNN detects approximately all the object regions from natural aphid images with various aphid distributions including dense aphid cliques and sparse aphid objects, in which an Improved Non-Maximum Suppression (INMS) strategy is proposed to eliminate overlapping regions. Then, the FCNN further refines the detected aphid regions from the CCNN. The final recognition and detection result would be obtained by combining the outputs from CCNN and FCNN together. Experiments on our dataset show that our CFN achieves an aphid detection performance of 76.8% Average Precision (AP), which improves 20.9%, 18%, 13.7% and 12.5% compared to four state-of-the-art approaches. Highlights: Domain specific dataset for aphid recognition and detection in the field. This dataset may have high application value in practical aphid monitoring. Coarse-to-fine network proposed for aphid detection in dense distribution regions. Improved non-maximum suppression can eliminate overlapping bounding boxes.