In response to the issue of estimating the recovery rate of residual film in cotton fields by current residual film recovery machines, a lightweight residual film recognition method named DCA-YOLO 11 was proposed, which enabled rapid and accurate identification of residual film on cotton field surfaces in natural environments. Taking the residual film on cotton field surfaces after the operation of the 4JMLE-210 residual film recovery machine as the research object, totally 900 images of residual film were collected at different time periods. Through preprocessing steps such as perspective transformation, image cropping, data cleaning, and data augmentation, a dataset of 5215 residual film sample images was constructed, which was divided into training and test sets at a 4∶1 ratio. To enhance the model’s performance, a depthwise convolution (DWConv) module was added to the backbone network of YOLO 11 to replace a standard convolution (Conv) module, thereby reducing computational complexity and the number of parameters. Additionally, a CBAM attention mechanism module was incorporated at the end of the detection output to improve the model’s perception capability and reduce interference from edges and backgrounds. Furthermore, the ADown module was used to replace the Conv module in the backbone network, enabling downsampling between different layers of the residual film feature maps, reducing the spatial dimensions of the feature maps while retaining key information to improve the accuracy of residual film target detection. Experimental results demonstrated that the DCA-YOLO 11 model achieved a precision (P) of 81.9%, a recall (R) of 80.9%, and a mean average precision (mAP) of 86.7% (at an IoU threshold of 0.5) in complex natural environments. The model has about 2.20 million parameters, and an FPS of 80f/s. Comparative experiments with other models showed that DCA-YOLO 11 outperformed YOLO v10, YOLO v9 and YOLO v8 in precision by 2.9 percentage points, 2.3 percentage points, 3.8 percentage points. In terms of recall, it was improved by 2.0 percentage points, 1.0 percentage points, and 1.8 percentage points compared with that of YOLO v10, YOLO v9, and YOLO v8, respectively. While its processing speed was slightly lower than than that of YOLO v10, and it surpassed YOLO v9 and YOLO v8 by 12.7% and 14.2%. DCA-YOLO 11 achieved the smallest model size and the fewest parameters while maintaining high accuracy, demonstrating its lightweight design and superiority. Through generalization test, the model’s detection results on the validation dataset showed an R2 of 0.72, a mean absolute error (MAE) of 4.92 pcs and a root mean square error (RMSE) of 2.72 pcs, indicating good generalization. The research result can provide a theoretical foundation and data support for the precise and efficient collection of residual film by recovery machinery in complex environments, as well as for the visual estimation of the recovery rate of residual film recovery machines.