Convolutional Networks for Real-Time Trademark Violation Identification

As an important part of the intellectual property rights protection, copyright infringement detection in the information age, is faced with threats monitored target for logos, images, social media platform trademarks. With the increase of user-generated process being located on the web, it is important to develop automatized systems allowing to identify trademarks violations in real time in order to ensure brand identity is preserved and that trademarks are not infringed upon; Traditional systems to detect trademarks are for the most part keyword-based systems or manual search, however this timeconsuming, has high-margin error and does not detect complex or slight modifications of the trademark. We tackle the problem of real-time detection of trademark violations and we propose a new approach based on Convolutional Neural Networks, a powerful class of deep learning algorithms that achieve high accuracy in image recognition problems. These imagegeneration and real-time features extraction step provided a base that guided for the newly proposed method which uses the power of CNNs in analyzing an image at every single timestamp for potentially matching images to learn the products that may have similarities to a flagged image for trademark violation. A deep CNN model was established and trained on a large dataset of branded logos and marks. Since the model has been trained to “see” images merely based on visual features (and not text), it can detect even minor differences in trademarks, whether that be color or distortion or shape. And this is especially important in cases of trademark infringement, where the infringer may try to take every effort to change the original trademark enough to evade getting caught. Our CNN model is built with several layers which learn more and more abstract features from images and is therefore fitting for this task as it is able to extract critical data to guarantee that the system is operating properly in an efficient manner. At the early layers, the model learns to identify simple features, like edges and textures, but eventually learns to correlate with complex features that may characterize the uniqueness of a trademark. One of the key advantages of CNNs in this type of applications is their ability to generalize — enabling the model to potentially detect violations in low-res images or noisy images, as it often the case with images processed in real-time. To accomplish real-time detection, the CNN model is docked to a streaming pipeline for dynamic images, and video processing. This means that even on multiple such as social media, ecommerce websites and more there will be constant vigilance of the content for trademark violations. It will provide brand owners with real-time alerts of in-progress infringements and enable swift enforcement actions such as takedown of infringing content and litigation.