Name
Title | ||
|---|---|---|
Recognition and counting of wheat mites in wheat fields by a three-step deep learning method |
Abstract | ||
|---|---|---|
The wheat mite always causes major damage in wheat plants and results in significant yield losses. Therefore, detecting wheat mites can provide important information, such as pest population dynamics and integrated pest management by monitoring wheat mite populations. However, the automatic classification and counting of wheat mites from images taken from crop fields are more difficult than those obtained under laboratory conditions, due to complicated background in crop fields, light instability and small wheat mites in images. Furthermore, the manual identification of wheat mites is very time-consuming and complex. Deep learning technique provides an efficiently automated way for address the issue. This paper proposes a three-step deep learning method to identify and count wheat mites from digital images. First, original large images are separated into smaller images as datasets. Then, the small images are labeled and then enlarged so that each of them can be located in corresponding position of original image. Second, one CNN takes an image (of any size) as input and outputs a set of feature maps for the image. Afterwards, the extracted feature maps are input to Region Proposal Network (RPN), which may be most likely the areas of wheat mites and output a set of rectangular objective proposals, each with an object score. Then one 256-d vector is generated from the obtained proposals by the other CNN. The vector is input into two fully connected layers, a box-regression layer and a box-classification layer, which output the probability scores of the position information and the population of wheat mites, respectively. Moreover, the superposition of the results for the small images is taken as the number of wheat mites for each original image. By using different backbone deep learning networks, ZFnet with five layers and VGG16 with sixteen layers achieved the accuracies of 94.6% and 96.4%, respectively. |
Year | DOI | Venue |
|---|---|---|
2021 | 10.1016/j.neucom.2020.07.140 | Neurocomputing |
Keywords | DocType | Volume |
Pest identification,Pest counting,Convolutional neural network,Region proposal network | Journal | 437 |
ISSN | Citations | PageRank |
0925-2312 | 0 | 0.34 |
References | Authors | |
0 | 9 | |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Chen Peng | 1 | 154 | 36.87 |
| WeiLu Li | 2 | 0 | 0.34 |
| SiJie Yao | 3 | 0 | 0.34 |
| Chun Ma | 4 | 0 | 0.34 |
| Jun Zhang | 5 | 0 | 0.34 |
| Bing Wang | 6 | 19 | 4.42 |
| Chun-hou Zheng | 7 | 732 | 71.79 |
| Chengjun Xie | 8 | 51 | 9.17 |
| Liang Dong | 9 | 326 | 52.32 |