Deep Learning Methods for Detecting Chilli Pests: A Novel Performance Analysis
More details
Hide details
School of Computer Science and Engineering, VIT-AP University, Amaravathi, India
Corresponding author
Kantha Raju Kanaparthi   

School of Computer Science and Engineering, VIT-AP University, Amaravathi, India
Ecol. Eng. Environ. Technol. 2024; 6:234-254
Ensuring food security is a top goal for all nations, yet infected plants can negatively impact agricultural production and the country's economic resources. In the past, farmers have depended on conventional techniques to enhance crop yield. In recent times, there has been a significant decline in crop production due to pest infestations on Chilli crops. The progress of deep learning techniques facilitates the categorization of diverse sorts of images in practical applications. Especially, detecting multi-class Chilli crop pests with good accuracy using deep learning algorithms is consistently a significant challenge. The proposed study concentrated in identifying pests on Chilli leaves using deep learning methods such as YOLOv5 and YOLOv7. To improve classification accuracy, a new and unique dataset called the standard balanced custom ‘Chilli pest dataset’ is created with 13,414 pest images. This dataset includes three specific pest classes: Black Thrips, Redmites, and White Fly. We analysed the custom Chilli dataset using YOLOv5 and YOLOv7 to evaluate their effectiveness in detecting pests in Chilli crops and obtained novel detection performance metrics. The resultant parameters mean Average Precision (mAP) for all three pest classes is 98.6% for YOLOv5 and 86.1% for YOLOv7. The YOLOv5s detector demonstrates superior performance compared to the YOLOv7 pest classification, with a 12.5% improvement. The YOLOv7 algorithm achieves its best classification accuracy (86.1%) at a lower epoch (110), while the YOLOv5 algorithm achieves its highest classification accuracy (98.6%) at a higher epoch (350). Nevertheless, despite this distinction, the YOLOv5 algorithm is recommended as the superior detector for accurately identifying pests in well-balanced multi-class pest type datasets, in comparison to YOLOv7, VGG-16 (~92.7%), and VGG-19 (~84.24%) deep learning architectures.
Journals System - logo
Scroll to top