Classifying and Detecting Live Insects with Computationally Effective Deep Learning Object Detection Models
Abstract A crucial part of agriculture is detecting insects that increase yield productivity. Insects in agricultural land are both helpful and harmful. The harmful insects are detected and controlled as early as possible, but these control measures should not affect the beneficial insects that help...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Springer
2025-06-01
|
| Series: | International Journal of Computational Intelligence Systems |
| Subjects: | |
| Online Access: | https://doi.org/10.1007/s44196-025-00885-6 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract A crucial part of agriculture is detecting insects that increase yield productivity. Insects in agricultural land are both helpful and harmful. The harmful insects are detected and controlled as early as possible, but these control measures should not affect the beneficial insects that help crops to grow. The existing pest detection models are image-based models where the preciseness of the insect detection is based on their appearance in the respective image, which may lead to the misclassification of insect classes if the insects are not present in the image properly. By analyzing consecutive frames rather than a still image, the proposed approach detects live insect objects from the video rather than a still image, where the presence of insects is identified by analyzing consecutive frames. As a result, insects can be detected without relying on the appearance of a single still image, which helps mitigate insects' misclassification. A wide range of applications in computer vision has proven that deep learning approaches are highly effective and popular. This study employs a variety of three deep learning-based object detection networks coupled with multiple backbone networks to maximize their efficiency. Each model is initially pre-trained using the COCO dataset to improve its performance. Experimental results show that SSD_MobileNet_V2 outperformed other models on insect classification and detection tasks. Regarding insect classification tasks, the SSD_MobileNet_V2 achieved an accuracy and F1 score of 98.02% and 97.99%, respectively. On the insect detection task, the mAP is 98.8% at a detection time of about 0.18 s. Also, it is delivered with a smaller model size of 6.5 MB, making it suitable for handheld devices. |
|---|---|
| ISSN: | 1875-6883 |