A Deep-Learning-Based Detection Method for Small Target Tomato Pests in Insect Traps

In a greenhouse environment where tomatoes are grown, pests in yellow sticky traps need to be detected in order to control the pest population. However, tomato pests typically found on yellow sticky traps are small in size and lack distinct visual features, making it difficult for convolutional networks to extract sufficient contextual information, thereby rendering the tasks of localization and classification exceptionally challenging. In this work, an improved approach based on the advanced object detection model You Only Look Once version 7-tiny (YOLOv7-tiny) is introduced, aiming to enhance the accuracy of detecting small tomato pests while maintaining computational complexity. Firstly, a context information extraction block (CIE) based on a Transformer encoder is proposed, and this block aims to capture global context, explore potential relationships between features, and emphasize important characteristics. Secondly, an Tiny-ELAN fusion network is introduced, which enhanced the feature fusion ability of the network. Thirdly, the feature fusion part takes the P2 feature layer into account and adds a P2 small target detection head. Finally, the SCYLLA-IoU (SIoU) loss function is introduced, and its components are redefined to incorporate direction information, which enhances the model’s learning ability and convergence performance. Experimental results show that our method can accurately detect three insects: whitefly (WF), macrolophus (MR), and nesidiocoris (NC) in the yellow sticky trap images of tomato crops. Compared with Faster R-CNN, SSD, YOLOv3-tiny, YOLOv5s, YOLOv7-tiny, YOLOv7, YOLOv7-x, YOLOv8n, YOLOv8s, YOLOv10n, and RT-DETR, the mean average precision of our method increased by 3.14%, 11.8%, 4.7%, 4.7%, 4.4%, 3.5%, 2.9%, 4.6%, 4.4%, 4.2%, and 4.2%, respectively.