Affiliations: [a] College of Agriculture, Hunan Agricultural University, Changsha, China
| [b] College of Information and Electronic Engineering, Hunan City University, YiYang, Hunan Province, China
| [c] Hunan Engineering Research Center for Internet of Animals, Changsha, China
| [d] Faculty of Computers and Information, Mansoura University, Mansoura, Egypt
Correspondence:
[*]
Corresponding author. Guan Chunyun, College of Agriculture, Hunan Agricultural University, Changsha 410128, China. E-mail: zhoulibo1@aliyun.com.
Abstract: A fundamental problem facing deep neural networks is that they require a large amount of data to keep the system efficient in complex applications. Promising results of this problem are made possible by using techniques such as data enhancement or transfer learning in large data sets. However, when the application provides limited or unbalanced data, the problem persists. In addition, the number of false positives generated by deep model training has a significant negative impact on system performance. This study aims to solve the problem of false positives and class imbalances by implementing an improved filter library framework for Cole pest identification. The system consists of three main units: First, the primary diagnostic unit (boundary box generator) generates a bounding box containing the location of the infected area and class. Then, the promising box belonging to each category is used as an input to the secondary diagnostic unit (CNN filter bank) for verification. In the second unit, the misclassified samples are filtered by training for each category of independent CNN classifiers. The result of the CNN filter bank is to determine if a target belongs to the category because it is detected (true) or no (false), otherwise. Finally, an integrated unit combines the information of the autonomous unit and the secondary unit in the future while maintaining a true positive sample and eliminating false positives of misclassification in the first unit. By this implementation, the recognition rate of this method is about 96%, which is 13% higher than our previous work in the complex task of Cole disease and pest identification. In addition, our system is able to handle false positives generated by bounding box generators and class imbalances that occur on data sets with limited data.
