<p>Small object detection has been widely applied in various fields, such as single-cell detection in biology, product defect detection in industry, and river garbage detection and localization from UAV perspectives. In commonly used object detection models, the shallow features extracted from the backbone network possess limited semantic information. This hinders their ability to effectively direct attention to core features. This limitation leads to a lower quality of features derived from the shallow layers of the backbone network, thereby constraining the feature fusion effectiveness of deep Feature Pyramid Networks (FPNs). To address this issue, we designed each layer of the deep FPN as a “teacher” to guide the corresponding layer of the backbone network in learning. Information from these layers is subsequently fused, and the fused deep FPN information is transferred to the shallow backbone network through non-invasive self-distillation, thereby enhancing the feature quality of the backbone network. This approach establishes virtual connections, forming our proposed Pseudo Feature Pyramid Network (P-FPN). Furthermore, we propose a two-stage Slicing Aided Hyper Inference (SAHI) detection algorithm based on a greedy strategy. This algorithm effectively addresses the problems of incomplete detection and multiple detections caused by “dirty data” (incomplete objects formed at slicing boundaries). We conducted comparative experiments on three public small object datasets across different scenarios and one private dataset. Compared with the baseline method, our proposed approach improved the average precision by 12% on the industrial PCB defect detection dataset. Additionally, when applied to the task of single-cell recognition in a microfluidic device, it achieved a 2% to 5% performance improvement in all five-fold cross-validations.</p>