LOGOS: A Brand-Independent Logo Detection Model
Logos are visual symbols that represent brands and companies. They are ubiquitous in the world around us, appearing on products, advertisements, websites, and more. Detecting and recognizing logos in images and videos has many useful applications, from measuring brand exposure to contextual advertising to intellectual property protection.
Most existing approaches to logo detection rely on recognizing specific brand logos, such as the iconic Apple logo or the golden arches of McDonald‘s. These models are trained on datasets with labels for each individual brand logo. However, this has some significant limitations:
-
Difficult to scale to large numbers of brands: There are millions of brands worldwide, with new ones being created all the time. Collecting sufficient training data for each one is impractical. Models struggle to recognize new or obscure brand logos they weren‘t trained on.
-
Class imbalance: Logo datasets tend to be dominated by a small number of highly popular brands. This can lead to poor performance on underrepresented brands.
-
Overkill for some use cases: Many applications only need to detect the presence and location of logos, not identify the specific brand. Current approaches are unnecessarily complex for these scenarios.
To overcome these challenges, we propose a brand-independent approach to logo detection. Rather than treating each brand as a separate class, we simplify the problem to just two classes: text logos and non-text logos. The key insight is that virtually all logos fall into one of these two categories. Text logos contain stylized brand names or slogans, while non-text logos are abstract symbols and shapes.
By framing logo detection as a binary text/non-text problem, we can leverage existing object detection architectures and datasets without the need for brand-specific labels. The training pipeline is as follows:
-
Take an existing large-scale logo dataset that has annotations for brand logo bounding boxes.
-
For each annotated logo, extract the image patch within the bounding box.
-
Run a text detection model on the extracted patches to determine if they contain text. We used the CRAFT text detection model for this.
-
Label each bounding box as either "text" or "non-text" based on the output of the text detector. Discard any ambiguous cases.
-
Train an object detection model using the re-annotated dataset. We used the RetinaNet architecture with a ResNet-50 backbone.
After training, the model is able to detect and localize logos in new images, outputting bounding boxes and confidence scores for the "text" and "non-text" classes. Some results are shown below:
[Good detection examples]As we can see, the model successfully detects a variety of both text and non-text logos across different domains. However, there are some failure cases:
[Bad detection examples]One issue is confusion with text that isn‘t part of a logo, such as signage or stylized text in ads. Separating true logo text from background text is challenging. There are also some missed detections and misclassifications between text and non-text.
Despite these limitations, the brand-independent approach has some key advantages:
- Simplicity: Only two classes to worry about, no need for huge taxonomies of brands
- Flexibility: The same model can be applied across industries and use cases
- Scalability: Training data is much easier to annotate and acquire
- Generalization: The model is more robust to new and unseen logos
In terms of applications, this simplified logo detector could be used for:
- Ad blocking and "debranding"
- Estimating overall brand presence and saturation in media
- Localizing logos for context-aware image captioning
- Filtering and organizing user generated content
- Training data collection for more fine-grained logo recognition
Future work could include adding more granular classifications beyond the text/non-text divide, such as typography, color, and shape attributes. Techniques for dealing with cluttered scenes, occlusion, and small logos are also important areas for improvement.
In conclusion, logo detection is a valuable capability with many applications. By stepping back from brand-specific recognition and reframing the problem around fundamental logo attributes, we can develop simpler, more scalable, and more generalizable models. The brand-independent approach represents a promising new direction for logo detection research and practice.
