Abstract: In order to solve the problems of large model calculation, difficult deployment and low recognition rate in cotton field weed detection, this study proposes a lightweight target detection model CEDL_YOLO based on the improvement of YOLO11n. The star block (Star) and the improved variant star block (VStar) are integrated in C3k2 to form the new C3k2_S and C3k2_VS, which prevents over-reliance on complex structures and hyperparameters, and strengthens feature extraction in scalable receptive fields. A channel mixing function (shift channel mix, SCM) is proposed to improve an efficient up-sampling module (efficient up-convolution block_shift channel mix, EUCB_S), eliminating feature redundancy after upsampling and enhancing the perception of details. A lightweight shared depth detection head (lightweight shared depthwise separable convolutional detection, LSDSCD) is proposed to reduce the model size and enhance multi-task collaboration and adaptability. DWConv (depthwise convolution) is introduced to replace the backbone network convolution to enhance the lightweight characteristics of the model. The results show that the improved CEDL_YOLO model has increased accuracy, recall, F1 score and mAP50 to 95.84%, 88.72%, 92.06% and 94.60% respectively. Compared with the original YOLO11n model, the floating point number, model size and parameter amount are reduced by 33.3%, 21.2% and 32.4% respectively, and the accuracy, recall, F1 score and mAP50 are increased by 0.45, 2.18, 1.47 and 1.05 percentage points respectively. This method meets the lightweight and accurate requirements of weed detection on mobile terminals, and provides a reference for cotton field weed detection and model deployment on mobile terminals.

Key words: deep learning, cotton field, weed detection, YOLO11n, lightweight

摘要： 针对棉田杂草检测中存在的模型计算量大、部署困难和识别率低问题，提出一种基于YOLO11n改进的轻量化目标检测模型CEDL_YOLO。在C3k2中集成星星模块（star block，Star）和改进的变异星星模块（variant star block，VStar）构成新的C3k2_S和C3k2_VS，防止过度依赖复杂结构和超参数，同时加强可扩展感受野中特征提取。提出通道混合功能（shift channel mix，SCM）改进了一种高效上采样模块（efficient up-convolution block_shift channel mix，EUCB_S），消除上采样后的特征冗余，增强对细节的感知。提出轻量化共享深度检测头（lightweight shared depthwise separable convolutional detection，LSDSCD），降低模型体积，同时增强多任务协同性和适应性。引入深度卷积（depthwise convolution，DWConv)替换主干网络卷积，增强模型的轻量化特性。结果表明，改进的CEDL_YOLO模型在准确率、召回率、F1分数和mAP50分别增加至95.84%、88.72%、92.06%和94.60%，与原YOLO11n模型对比，浮点数、模型大小和参数量分别减少33.3%、21.2%和32.4%，准确率、召回率、F1分数和mAP50分别提升0.45、2.18、1.47和1.05个百分点。该方法满足杂草检测在移动端部署的轻量化和精准要求，为棉田杂草检测及模型在移动端的部署提供参考。

关键词: 深度学习, 棉田, 杂草检测, YOLO11n, 轻量化