Abstract: In modern precision agriculture, rapid and reliable rice pest detection is essential for reducing yield loss, supporting timely field management, and decreasing excessive chemical pesticide application. However, automatic pest detection in paddy fields remains challenging because rice pests usually appear as small targets with weak texture, limited pixel occupancy, irregular posture, and dense spatial distribution. In real field images, dynamic natural illumination, overlapping rice leaves, water-surface reflection, motion blur, and background textures similar to pest bodies often cause missed detections, false detections, and localization deviation. Existing agricultural object detection models have achieved promising progress, but many methods improve accuracy by introducing complex attention mechanisms or heavy detection heads, which may increase computational cost and limit deployment on edge devices. Lightweight models, on the other hand, may lose fine-grained target features during downsampling, resulting in insufficient representation of small rice pests. To address these problems, this study proposes a lightweight small-target rice pest detection model named YOLOv11n-BIP, where BIP denotes a Balanced Information-preserving Pyramid designed to coordinate feature extraction, information-preserving downsampling, and pre-detection feature recalibration.The proposed YOLOv11n-BIP model is constructed on the YOLOv11n baseline and optimized from three complementary aspects. First, a Cross Stage Partial Heterogeneous convolution module (CSPHet) is introduced into the backbone and neck feature extraction units. By combining heterogeneous convolution kernels with cross-stage information fusion, CSPHet enhances multi-scale texture representation while maintaining a compact structure, which helps the network distinguish small pest bodies from rice leaves and complex field backgrounds. Second, an ADown information-preserving downsampling path is designed for the bottom-up feature aggregation process. Compared with standard stride-2 convolution or pooling, ADown uses parallel convolution and pooling branches to retain local edge responses and salient texture information during spatial resolution reduction, thereby reducing the risk that small pest features disappear in deeper feature layers. Third, an Efficient Multi-scale Attention (EMA) module is embedded before the detection heads at the P3, P4, and P5 scales. EMA does not replace the prediction head or participate in final decoding; instead, it recalibrates the neck output features before classification and regression, suppressing non-target activations caused by water reflection, leaf occlusion, and background noise while strengthening responses in pest regions.Experiments were conducted on a rice pest image dataset containing 4 553 images and 11 pest categories, with the data divided into training, validation, and test sets by stratified random sampling at an approximate ratio of 7:2:1. All compared models were trained under the same input resolution, training epochs, optimization settings, and data augmentation strategies to ensure fairness. The experimental results show that YOLOv11n-BIP achieves better detection accuracy while preserving lightweight characteristics. Compared with the original YOLOv11n baseline, the proposed model improves precision by 2.9 percentage points and mAP@0.5 by 3.1 percentage points, increasing mAP@0.5 from 76.8% to 79.9%. The model contains only 2.604 M parameters and reaches 46.8 FPS under the same inference environment. Ablation experiments further demonstrate that CSPHet, ADown, and EMA contribute complementary improvements: CSPHet strengthens fine-grained feature extraction, ADown improves information retention during feature transmission, and EMA enhances robustness under complex backgrounds. Confusion matrix analysis also indicates improved recognition for small targets and relatively scarce categories, such as Phlaeothripidae, Chloropidae, and Ephydridae.In summary, YOLOv11n-BIP achieves a favorable balance between detection accuracy, model size, and inference efficiency. The model provides an effective technical solution for small-target rice pest detection in complex paddy-field scenarios and offers practical support for intelligent pest monitoring, edge deployment, and precision pest control in smart agriculture.