Cotton boll count is a critical phenotypic trait for estimating cotton yield and plays a vital role in precision agricultural management. However, accurately detecting cotton bolls in densely planted fields remained challenging due to complex backgrounds, occlusion, and varying illumination conditions. High-resolution UAV imagery was employed to capture cotton field scenes in a densely planted area of Xinjiang. A comprehensive dataset was developed through image segmentation and augmentation techniques, ensuring diverse representations of field conditions. To address the trade-off between detection accuracy and computational efficiency, an improved lightweight detection model IML-YOLO was proposed. The model integrated a novel GRGCE module that combined efficient ghost convolution with a RepGhostCSPELAN structure for feature extraction, a CAHSFPN feature fusion mechanism to enhance multi-scale representation, and a Focaler-MPDIoU loss function to refine localization accuracy. Extensive experiments demonstrated that IML-YOLO reduced computational complexity by 32.1%, decreased model size by 47.5%, and lowered parameter count by 50% compared with that of the baseline YOLO v8n, while boosting mean average precision by 10.1 percentage points. Furthermore, when applied to cotton yield prediction, the model achieved an average relative error of only 7.22%. These findings indicated that the proposed IML-YOLO model and yield prediction methodology can offer an effective solution for real-time cotton boll detection and significantly contribute to the advancement of intelligent cotton management.