Inappropriate image detection is a key component of any content moderation system. It identifies images that portray explicit nudity, violence or are visibly disturbing.
A total of 40 journals containing the search term “western blot” were screened for inappropriate duplications of figures. The journal impact factor varied widely, ranging from the Journal of Cell Biology (0.3%) to the International Journal of Oncology (12.4%).
Content Moderation
With an endless stream of user-generated content flowing through platforms, brands need a way to keep track and remove harmful content. Moderation software helps protect users from harm, safeguard a company’s reputation, and ensure legal compliance.
Moderation software is a vital tool for any platform or website that allows users to post text, images, or video. It filters out content that is inappropriate, offensive, or illegal, protecting the brand from potentially damaging legal ramifications.
Many companies use different methods of moderation, such as keyword filtering or image recognition. Keyword filtering uses algorithms to identify words or phrases that are considered inappropriate for a platform and can be prone to false positives (for example, if AI misunderstands sarcasm). Image recognition is more accurate for detecting overtly offensive images such as nudity, hate speech, violence, drugs, and offensive gestures. However, human moderators can be a better fit for identifying the more subtle aspects of harmful content such as propaganda and disinformation.
NSFW Detection
NSFW (Not Safe for Work) is a shorthand warning label that essentially says “this content might be inappropriate to view in the workplace.” Initially, it was used mostly to reference sexual content or nudity. It has since evolved to include a range of delicate topics that might trigger the viewer, such as gory violence or distressing imagery.
Whether you’re dealing with user-generated content or professional images, it can be easy for rogue users to tarnish your brand image with nude or inappropriate content that may not be suitable for viewing in public places or at work. With image recognition, you can keep NSFW images out of your content and prevent the negative reactions that can hurt your business.
Image recognition is a complex task that requires the use of advanced algorithms to identify and categorize the content. The model shared in this article is trained with one type of NSFW content: pornographic images. Defining what is considered NSFW is highly subjective, and what may be objectionable to one person could be perfectly acceptable in another context.
Violation Detection
Detecting vehicles that violate traffic rules is a challenging task for law enforcement and traffic management. The process involves analyzing CCTV footage and recognizing vehicles. This method is laborious since each vehicle has to be tracked without a miss. To automate this process, it is possible to use machine learning to recognize vehicles and alert authorities and violators of the violation. The system utilizes YOLOv3 (You Only Look Once version 3), which is a convolutional neural network, and Darknet-53 as a feature extractor. Once a vehicle is detected, the cropped image of the license plate is outputted and displayed with its corresponding type of violation.
Reporting
Inappropriate image detection detects content in messages that could be harmful to your organization, such as explicit adult content or violent images. When these images are flagged, they’re available in the moderation dashboard for your team to review and take appropriate action.
A two-step dataset documentation process helps users identify potentially inappropriate images in pre-trained models. Prompt-tuning based on a dataset of socio-moral values steers the CLIP model to classify inappropriate image content, reducing manual human effort. A second step documents this subset with word clouds based on captions generated using a vision-language model.
These word clouds highlight the most frequent concepts for documentation purposes. Among them are several National Socialist symbols especially the swastika and persons in Ku-Klux-Klan uniform, insults including sex, drugs and weapons (e.g. a pistol or knife) as well as the depiction of naked bodies. Additionally, the word cloud identifies images that might be considered sexually disturbing (e.g. a bathtub tainted with blood). In addition to the inappropriate concepts detected, the classifier also reports on potentially sensitive information types such as financial data.