In order to achieve precise picking of residual film in farmland, improve the recovery rate of residual film recovery machine. The improved Faster R-CNN convolutional neural network is applied to the identification and detection of residual film in farmland, and a method of identifying residual film in farmland is proposed. Taking the residual film left on the surface of the farmland after the 11MS-1850 residual film recovery machine worked as the research object, a total of 1648 images were collected during different periods of sunny and cloudy days. Expand the data set by changing the image brightness, rotation, etc., and finally get 4950 residual film sample images, which are divided into training set (3465), validation set (990), and test set (495) according to 7:2:1; The dual-threshold algorithm is used to replace the traditional single-threshold algorithm, which reduces the impact of thresholds on model performance; through comparative experiments, ResNet50 with a residual network structure is selected as the backbone feature extraction network. The accuracy rate can reach 88.84%, and the recall rate is 87.70%. , The overall accuracy is 88.27%; in order to make the detection model more sensitive to small targets, according to the size of the residual film in the data set, the scale parameters of 322 and 642 are added to the original anchor points, and the accuracy, recall, and overall accuracy have been improved. 1.29, 0.67, 0.97 percentage points, the single detection time is 284.13ms, which basically meets the requirements for identifying residual film. It can provide a reference for the installation of replenishment equipment for the residual film recovery machine, and provide a theoretical basis for the development of artificial intelligence residual film recovery machine.