Selecting logo images using machine-learning-logo classifiers

#Software developers and engineers have designed logo-detection systems to automatically detect logos within digital images and identify positions of such logos within the digital images. Detected logos may correspond to different logo classes, such as a logo for a specific automobile company and a logo for a specific clothing company. In practice, logo-detection systems have been applied to recognize vehicles for intelligent traffic control, recognize logos to address consumer confusion or perception, or detect use of a trademark or trade dress. To train a machine-learning model to detect logos and corresponding logo positions, some logo-detection systems use synthetic images or weakly labelled real images lacking ground-truth data for logos within images. Despite significant advances in logo detection, computing limitations inhibit existing logo-detection systems from accurately detecting logos in real images, developing real images with corresponding ground-truth data, and scaling a sample of real logo images to train machine-learning models. Some existing logo-detection systems also apply training algorithms that foment model drift during the training of machine-learning models.

#In some cases, for example, conventional closed-set-logo-detection systems cannot develop (or train on) sufficient samples of real logo images with corresponding ground truths to accurately detect logos of certain logo classes. Because annotating real-world images that display logos can be time consuming and computationally expensive, some conventional techniques cannot produce, in practice, a critical volume of such images to accurately train a logo-detection system. Rather than rely on real-world images for training, some existing logo-detection systems train on synthetic images and consequently often fail to train the system to consistently classify logos within real-world images with accuracy. Indeed, certain logos lack sufficient corresponding real-world images to train a machine-learning model. Because such annotation is impractically cumbersome and real-world images of certain logos are sparse, some existing logo-detection systems also cannot train a machine-learning model to accurately detect or identify a position of a logo for certain logo classes.

#In addition to inaccurate detection, some existing logo-detection systems inaccurately train a machine-learning model by failing to detect or correct model drift. When a machine-learning model incorrectly detects a logo within an input image—by, for example, falsely detecting a logo or failing to detect the presence of the logo—the logo-detection system can adjust or maintain internal parameters of the machine-learning model to reinforce such errors. By propagating errors in logo prediction through training iterations, the machine-learning model can drift to configure parameters to repeat such errors in future logo classifications.

#The detailed description refers to the drawings briefly described below.

#FIG. 1 illustrates a logo-detection-training system training a machine-learning-logo classifier based on synthetic training images and incrementally replacing the synthetic training images with logo images in accordance with one or more embodiments.

#FIG. 2 illustrates the logo-detection-training system generating synthetic training images in accordance with one or more embodiments.

#This disclosure describes one or more embodiments of a logo-detection-training system that selects logo images comprising ground-truth logos corresponding to a logo class by initially training a machine-learning-logo classifier using synthetic training images and iteratively applying the machine-learning-logo classifier to identify logo images as replacements for the synthetic training images. To identify logo images and corresponding annotations for ground-truth-logo positions, the logo-detection-training system can apply the machine-learning-logo classifier to determine logo scores indicating predicted logos within candidate logo images and boundary identifiers for predicted logos within the candidate logo images. By augmenting the iterative training of a machine-learning-logo classifier to include user curation and removal of incorrectly detected logos from candidate logo images via a curation user interface, the disclosed logo-detection-training system can further avoid model drift across training iterations.

#In some embodiments, for instance, the logo-detection-training system generates a set of synthetic training images portraying logos from a particular logo class and trains a machine-learning-logo classifier based on the synthetic training images. The logo-detection-training system can further apply the machine-learning-logo classifier to candidate logo images to select a subset of logo images. Upon identifying the subset of logo images, the logo-detection-training system can replace a subset of the synthetic training images with the subset of logo images to generate a set of mixed training images. Based on the set of mixed training images, the logo-detection-training system can retrain the machine-learning-logo classifier. By incrementally replacing the synthetic training images with selected candidate logo images, the logo-detection-training system can select logo images comprising ground-truth logos corresponding to the logo class using the retrained machine-learning-logo classifier.

#As just mentioned, in some embodiments, the logo-detection-training system generates and utilizes synthetic training images to initially train a machine-learning-logo classifier. For example, the logo-detection-training system can generate synthetic training images by inserting a logo from a logo-class as an overlay to an existing digital image. The logo-detection-training system can select a size, orientation, and location of the logo within the digital image and then apply the logo to generate a synthetic training image. Because the boundary of the logo within the synthetic training image is known, the logo-detection-training system can then train the machine-learning-logo classifier based on the synthetic training images and known boundary identifiers.