Focused on the issues of high miss rates and localization deviations in standard YOLOv11 caused by small, diverse, and occluded targets like irregularly shaped packaging on warehouse shelves, an improved algorithm IYOLOv11 based on the collaborative optimization of global perception and lightweight convolution was proposed. Firstly, a Context-aware Global Block (CGBlock) and a Wavelet Pyramid Convolution Network (WPCN) were embedded into the backbone to enhance contour feature extraction of occluded targets via context attention and frequency-domain transformation mechanisms. Secondly, the Ghost Shuffle Convolution (GSConv) module was employed to reconstruct a lightweight feature fusion layer in the neck, and an optimized fusion loss function was introduced to further improve bounding box localization accuracy and multi-scale feature fusion efficiency. Experimental results on the SKU-110K dataset showed that with a marginal parameter increase of only 0.695 M, IYOLOv11 achieved a 91.94% Mean Average Precision mAP50, which was 4.59 percentage points higher than the baseline. Moreover, the high-threshold precision mAP50:95 improved by 5.08 percentage points, along with a 12% increase in inference speed. Compared to other mainstream lightweight detection networks, the proposed algorithm demonstrates stronger robustness when processing densely occluded and complex-shaped targets, satisfying the practical needs of warehouse automation systems for high-precision detection while maintaining a high inference speed.