Object detection in complex scenes is frequently challenged by the low visibility of small objects and frequent occlusions,leading to missed detections and compromised robustness. To overcome these limitations, this paper introduces YOLO-MB, an enhanced model that integrates three key components: a Multi-scale Dual-path Enhancement (MSDE) module, aBidirectional Feature Pyramid Network (BiFPN), and a Dynamic ReLU (DY-ReLU) activation mechanism. The MSDEmodule, embedded in the neck of the network, employs multi-scale convolutional branches and a dual attention mechanism to intensify feature representation for small and occluded objects. BiFPN supplants the original feature fusion schema tofacilitate efficient multi-scale feature interaction and information exchange. Furthermore, DY-ReLU is introduced into the detection head to adaptively generate piecewise activation parameters based on global contextual information, enhancing feature discriminability without increasing inference parameters or computational cost. The synergy of these three modules forms a comprehensive optimization pipeline from feature enhancement to fusion and adaptive activation, significantly improving detection performance in occlusion-dense scenarios. Extensive evaluations on the public COCO dataset and a self-built CTOCDD dataset demonstrate that YOLO-MB surpasses existing YOLO variants and other leading methods inboth accuracy and recall, achieving 43.2% mAP50-95 on COCO with only 2.4M parameters and 1.46ms inference latency, confirming its superior performance and efficiency.