Fig 1.
(A) A pest with different shapes and sizes. (B) Various rice pests with different shapes and sizes in the IP102 dataset. (C) Tiny rice pests in the field in the AgriPest dataset. (D) Not obvious rice pests.
Fig 2.
The structure of CATransU-Net.
Fig 3.
The main components of CATransU-Net.
(A) Residual dilated Inception (RDI) module in encoder. (B) Deconvolution in decoder. (C) CASC. (D) DTA.
Table 1.
The image distribution of rice pest image subset.
Fig 4.
Rice pest image examples and Class distribution.
(A) Simple pest images. (B) Complex pest images. (C) The number of pest images of different categories. (D) 14 augmented images of the first pest image.
Fig 5.
The losses of mAP of four models with the number of iterations.
Fig 6.
The detected pest images of obvious pests by 7 methods.
(A) Original simple rice pest images and labeled pest images. (B) U-Net. (C) TransUnet. (D) DMSAU-Net. (E) HDLU-Net. (F) Swin Transformer. (G) TinySegformer. (H) CATransU-Net.
Fig 7.
The detected pest images with complex background by 7 methods.
(A) Original complex rice pest images and labeled pest images. (B) U-Net. (C) TransUnet. (D) DMSAU-Net. (E) HDLU-Net. (F) Swin Transformer. (G) TinySegformer. (H) CATransU-Net.
Table 2.
The detected results of 6 methods and their training time.
Table 3.
Pest detection accuracies by adding different modules.
Fig 8.
The comparison of detected pests by adding each component.
(A) Original images. (B) U-Net. (C) U-Net+ RDI. (D) U-Net+ DTA. (E) U-Net+ CASC.
Fig 9.
The comparison of detected pests by adding two components.
(F) U-Net + DTA and CASC. (G) U-Net+ RDI and DTA.
Table 4.
Pest detection set and results.
Fig 10.
Visualized detected pest images with complex shapes and background by 7 methods, where (A) Original complex pest images and labeled pest images, (B) U-Net, (C) TransUnet, (D) DMSAU-Net, (E) HDLU-Net, (F) Swin Transformer, (G) TinySegformer and (H) CATransU-Net.