In this paper, a semi-supervised regression method based on a Generative adversarial network is proposed to meet the demand for online sorting for waste textile recycling, which uses some labeled samples and a large number of unlabeled samples to train the semi-supervised regression. The semi-supervised regression consists of a generator composed of neural networks and a discriminator composed of neural networks. The generator generates a mixed sample that is as close as possible to the actual labeled and unlabeled training dataset content. The discriminator is used to validate the samples generated by the generator and predict the continuous labeling of these samples. The generated network is trained through feature matching loss, which is the average error between the output of the actual sample in the middle layer of the discriminator and the generated sample. The discriminant has two outputs, one for predicting sequence markers and the other for determining the probability of whether the generated sample is a true or false sample. Train discriminants by using a combination of traditional unsupervised generative adversarial network loss functions and supervised regression losses. The generated network is trained through feature matching loss, which is the average error between the output of the actual sample in the middle layer of the discriminator and the generated sample. We have collected 400 blended samples with different wool contents and 3 000 blended samples with unknown components. 70% of labeled and unlabeled mixed samples were randomly selected as the training set, while the remaining 30% of labeled samples were used as the test set for repeated experiments. This article has conducted multiple experiments for verification. The first experiment is a blended spectrum generation experiment, which is used to verify that the generative adversarial network can effectively generate mixed sample spectra based on intrinsic laws. The second experiment is a semi-supervised adversarial network quantitative analysis performance comparison experiment, which trains and tests the wool composition analysis model, and compares the performance of this semi-supervised adversarial network quantitative analysis model with other quantitative models. The third experiment is a prediction and comparison experiment of on-site high wool content blended yarn segmentation models. Composition analysis is conducted on blended yarn samples with wool content between 80% and 99%, and the performance of this semi-supervised adversarial network quantitative analysis model is compared with that of other quantitative models. The fourth is a field prediction experiment for the subdivision comprehensive model of medium to high wool content blended fabrics. A semi-supervised adversarial network quantitative analysis model is trained using blended fabric samples with wool content between 40% and 99% and deployed in the sorting system. The operator conducts on-site testing data for accuracy, analysis time, and other tests. The experimental results show that the semi-supervised regression method based on a Generative adversarial network is superior to PCR，PLSR, SVR, BPNN and other models, and the prediction R2 of this model reaches 0. 94. After repeated on-site testing, the model can quickly extract blended samples with a wool content of over 40%.