The new strategy will effectively improve the accuracy of agricultural pest identification

Recently, the research paper completed by Qian Wanqiang, a researcher at the Institute of Agricultural Genomics of the Chinese Academy of Agricultural Sciences (Shenzhen), was published in the Journal of Agricultural Sciences (English) (Journal of Integrative Agriculture,JIA) officially published. This study proposes a two-stage segmentation strategy, which can effectively deal with the segmentation of small targets in complex backgrounds, and the segmentation effect is significantly improved compared with the single-stage model, which can provide high-quality image data for pest larval identification and provide an effective reference for the segmentation of small targets in complex backgrounds represented by pests.

Many forms of field pests. Photo courtesy of Institute of Genomics, Academy of Agricultural Sciences

Online automatic identification of field pests is an important auxiliary means for farmland pest control. In practical applications, due to small targets, high species similarity, complex background and other factors, the accuracy of the insect online identification system is seriously affected. Therefore, before image classification, segmentation of the detection target as completely as possible is an effective means to improve the accuracy of image recognition.

Based on the analysis of farmland pest control background, an image semantic segmentation dataset containing nine common pests was established, and four data augmentation methods were used to balance the number of simple background and complex background samples in the dataset. A lightweight semantic segmentation algorithm is improved to meet the requirements of mobile applications in real scenarios. A two-stage pest segmentation network framework was designed, and its effectiveness was verified by targeted experiments.

The proposed two-stage semantic segmentation algorithm MRUNet borrows the practice of Mask R-CNN for object detection before semantic segmentation in terms of structure. In terms of specific operations, MRUNet built a Faster R-CNN object detection model for larval target detection. Targeted adjustment of the position information obtained by object detection and used to intercept sample image fragments in the original high-resolution image; Finally, the obtained image fragments mainly containing pest larvae were semantically segmented using the lightweight UNet algorithm based on improved MobileNet. In quantitative experiments, qualitative analysis and statistical analysis, the proposed MRUNet algorithm outperforms other algorithms in terms of segmentation accuracy and model stability. (Source: Li Chen, China Science News)

Related paper information:

Source link

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *

Back to top button