SOE-YOLO: A Small-Object Enhancement Detection Network

With the development of deep convolutional neural networks, significant progress has been made in object detection over the past few years. Current detection algorithms demonstrate high accuracy for large- and medium-sized objects. However, the detection of small objects remains a significant challenge due to their limited pixel area, scarcity of distinctive features, low contrast against complex backgrounds, and increased sensitivity to boundary changes. These limitations reduce both accuracy and inference speed in small-object detection tasks. To address these challenges, we propose a small-object enhancement (SOE) network based on YOLOv8, named SOE-YOLO. The network is designed to enhance both feature extraction and semantic understanding of small objects. SOE-YOLO integrates two key modules: the attention-guided multipath spatial pyramid pooling (AMSPP) module and the cascaded semantic embedding (CSE) module. The AMSPP module initially extracts multiscale information through multipath spatial pyramid pooling. To guide the network in focusing on the features of smaller objects, we introduce the marginal attention (MA) mechanism, which efficiently and accurately assigns corresponding weights to each feature point. Subsequently, the CSE module stacks multiple semantic embedding units (SEUs), progressively incorporating higher-level semantic information into high-resolution feature maps. This strategy effectively combines the spatial details of small objects with abstract, high-level semantic knowledge, enhancing the model’s ability to discern complex relationships between small objects and backgrounds. Extensive experiments on benchmark datasets demonstrate the effectiveness of SOE-YOLO, which outperforms other state-of-the-art detectors in terms of both detection accuracy and robustness for small-object detection.