Small target detection is of great significance in the field of computer vision, and is widely used in scenarios such as autonomous driving and intelligent monitoring, but due to the problems of small target size and insufficient pixel information, the existing detection algorithms perform poorly in occlusion scenarios. Small targets themselves contain limited features, and occlusion further reduces the visible area and increases the difficulty of detection. Most of the existing detection frameworks are based on global characteristics and cannot effectively focus on the details of small targets. In addition, small target detection is often sensitive to resolution and scale changes, and occlusion leads to partial or complete coverage of target features, which affects the positioning and classification accuracy of the model. In dense scenes, occlusion and overlap between multiple small targets aggravate the difficulty of detection and seriously affect the detection effect. In this paper, we propose an occlusive small target detection method that combines global feature extraction and local detail enhancement. Specifically, on the basis of deep convolutional neural network, a region self-attention mechanism is introduced to reduce the influence of background noise on detection by adaptively paying attention to the occluded small target area. At the same time, a feature fusion strategy based on multi-scale feature pyramid is proposed, which enhances the network’s perception ability of small targets and improves the detection accuracy of small targets through cross-layer feature fusion. In addition, a data augmentation method based on Generative Adversarial Networks (GAN) is introduced to improve the robustness of the model to complex occlusion scenarios by generating synthetic data with different occlusion cases.