As disclosed, f-scores can be generated for apparel items. Training images are identified, where each training image is associated with a corresponding set of tags including information about a plurality of attributes. A first convolutional neural network (CNN) is trained based on the plurality of training images and a first attribute. The first CNN is iteratively refined by, for each respective attribute, removing a set of neurons from the first CNN and retraining the first CNN based on the training images and the respective attribute. Upon determining that the first CNN has been trained based on each of the attributes, one or more CNNs are generated based on the first CNN. An image is received, where the image depicts an apparel item. The image is processed using the one or more CNNs, and an f-score for the apparel item is determined based on the output.