A raw cotton impurity image segmentation method based on improved UNet model

Abstract

To improve the segmentation accuracy and timeliness of raw cotton mixed impurity images, a new improved UNet model of raw cotton impurity segmentation algorithm was proposed. A new coding module was designed based on the ResNet50 structure, and one convolutional layer was used to replace the last three residual modules and the fully connected layer to implement optimization of the model parametric quantities and improve the extraction of image feature information. The loss function of CEloss and Dice loss was designed to optimize the imbalance between positive and negative samples and improve the segmentation ability of the model for fine impurities. Finally, the migration learning method was applied to pre-train the encoder for the initialization of the backbone network weights based on the VOC dataset to further optimize the slow convergence of the model due to the small amount of image data. The experimental validation shows that the MPA value of the method is improved by 24.52%, the accuracy rate reaches 88.24%, and the processing time of a single image is shortened to 21.9 ms, which is 49.42% higher than that of the traditional method. The improved method described is effective in improving the accuracy and timeliness of image segmentation of raw cotton impurities.

Keywords

• residual network
• unet
• transfer learning
• raw cotton impurities
• image segmentation