<p>Aiming at the problems of small-scale clothing miss detection and false detection, unclear boundary segmentation caused by multi-scale clothing images, a novel clothing instance segmentation algorithm based on improved Mask R-CNN is proposed. Firstly, the Bi-directional Feature Pyramid Network is used to optimize the feature extraction network, balance the feature extraction ability of the model for different sizes of clothing in the image, and improve the segmentation performance of small-scale clothing. Then, dilated convolution is used to improve the ResNet network structure to prevent small-scale clothing from losing the feature information due to the dimensionality reduction operation while increasing receptive field. Finally, the regression loss function is optimized by using the Log-Cosh function to reduce the impact of error values on the model. The experimental results show that compared with the original Mask R-CNN network model, the proposed method improves the segmentation accuracy of small-scale clothing by 2.5%, the average precision of the proposed method is improved by 5.7%.</p>