Aiming at the problems of complex image background, high miss detection rate, large number of model parameters and long training time in cocoon image defect recognition, this paper proposes a cocoon defect detection module YOLO11-BS that improves YOLO11s. The model replaces the convolution block in the backbone network with a BSNet module, which adopts depth-separable convolution and width multiplier optimization, reducing the number of parameters and computation volume. It reduces the number of model parameters and computation and realizes a lightweight design. Meanwhile, the SCAM module is embedded in the neck network, which enhances the feature extraction ability of the model for complex backgrounds and small targets through spatial and channel attention mechanisms. In this paper, a manually labeled cocoon dataset with data augmentation is used to train and perform model evaluation and ablation experiments on a RTX4060ti platform. The experiments show that the improved YOLO11-BS model outperforms the original YOLO11s model in terms of in-precision (P), recall (R), and average precision (mAP50). The ablation experiments validate the effectiveness of the BSNet and SCAM modules, demonstrating their significant role in improving model performance and robustness. This study provides an efficient solution for cocoon target detection task.