Fashion image analysis has a lot of trouble working with noisy data, especially when there aren't any good training examples to use. This happens because standard auto encoders can't keep class separability when the noise level changes. This paper introduces Structured Auto encoders (SAE), which combine convolutional encoder-decoder architectures with geometric constraints in the latent space. This strategy uses greedy layer-wise pre-training, and then it fine-tunes the model using two loss functions: structured latent space regularization and reconstruction error. The encoder has convolutional layers with 3×3 filters and ReLU functions that only turn on when they are needed. During testing, it was tested on Fashion-MNIST with noise levels ranging from 20% to 70%. It was also tested on MNIST, DeepFashion2, and 3D human pose datasets. The results of the experiment show that SAE can classify 85.4% of the time with only five samples labeled for each group. This is much better than the standard auto encoders (68.4%) and the baseline support vector machine (SVM) methods (84.32%). With the best setup, which has 1,000 hidden nodes and a learning rate of 0.1, images with up to 70% noise can be reconstructed well With 80% classification accuracy on clean test data, the latent space structured method allows for the solution of basic issues that are related to introducing geometric constraints between the classes so that the class boundaries remain intact even in a noisy environment.The proposed method achieves a classification accuracy of 85.4% ± 1.2%under standard evaluation settings, with robustness validated under noise levels of up to 70%. All results are obtained using standard training–testing splits and are averaged over multiple experimental runs to ensure reliability and reproducibility. The proposed semi-supervised performance is enhanced by increasing the number of discriminative features through the use of a highly structured representation of the data, as opposed to using an unstructured representation, which results in performance enhancement. The author presents an original framework capable of simultaneously providing robust image classification and denoising capability through geometric constraints established in the framework, which enhances the learning of discriminative features through the use of limited amounts of labeled data. The implementation of this framework in the real-world fashion industry would facilitate the processing of fashion images with respect to noise resistance and the ability to annotate images quickly.