Rice panicle detection is a key technology for improving rice yield and agricultural management levels. Traditional manual counting methods are labor-intensive and inefficient, making them unsuitable for large-scale farmlands. This paper proposes FRPNet, a novel lightweight convolutional neural network optimized for multi-altitude rice panicle detection in UAV images. The architecture integrates three core innovations: a CSP-ScConv backbone with self-calibrating convolutions for efficient multi-scale feature extraction; a Feature Pyramid Shared Convolution (FPSC) module that replaces pooling with multi-branch dilated convolutions to preserve fine-grained spatial information; and a Dynamic Bidirectional Feature Pyramid Network (DynamicBiFPN) employing input-adaptive kernels to optimize cross-scale feature fusion. The model was trained and evaluated on the open-access Dense Rice Panicle Detection (DRPD) dataset, which comprises UAV images captured at 7 m, 12 m, and 20 m altitudes. Experimental results demonstrate that our method significantly outperforms existing advanced models, achieving an AP50 of 0.8931 and an F2 score of 0.8377 on the test set. While ensuring model accuracy, the parameters of the proposed model decreased by 42.87% and the GFLOPs by 48.95% compared to Panicle-AI. Grad-CAM visualizations reveal that FRPNet exhibits superior background noise suppression in 20 m altitude images compared to mainstream models. This work establishes an accuracy-efficiency balanced solution for UAV-based field phenotyping.

Thrips are major minute pests that threaten crop growth, and their early and accurate detection is of great significance for pest monitoring and precision control. Existing image-based thrips detection studies have mainly focused on insects captured on sticky traps, whereas direct detection of thrips on plant surfaces under complex leaf-background conditions remains relatively underexplored. Moreover, in leaf-background images, thrips are extremely small, have indistinct contours, and are susceptible to interference from noise such as leaf textures and spots, resulting in insufficient detection and counting accuracy. To address these issues, this study proposes SCIC-DEIM, a small-object thrips detection model based on the end-to-end detector DEIM. First, leaf-background thrips images were collected at Jingchu University of Technology to construct a dataset containing 5618 images and 47,726 thrips instances. Second, focusing on three key aspects—background redundancy suppression, target response enhancement, and detail information recovery—the SCIC-DEIM model was developed. Specifically, the SCConv and IIA modules were introduced into the backbone network to construct SCIStage, and CARAFE was adopted to replace conventional upsampling operations. Experimental results show that, compared with DEIM, SCIC-DEIM improves Precision, mAP@50, and mAP@50:95 by 2.4, 2.0, and 2.9 percentage points, respectively, while reducing GFLOPs by approximately 20.6%. It also achieves improved MAE, RMSE, and R² values compared with DEIM. The proposed algorithm demonstrates strong capability in small-object thrips detection under leaf-background conditions and can provide a methodological reference for intelligent monitoring of tiny pests on leaf surfaces in protected agriculture.