AI nudity detection

AI nudity detection overview

The AI nudity detection feature identifies explicit content in videos. It's offered as part of Gcore's AI Video Services. AI nudity detection uses machine learning to provide automatic visual content analysis, recognition, and moderation.

Gcore's AI nudity detection can be used for moderating user-generated content (UGC). It can help you to determine whether a video is suitable for all user categories without age restrictions, or if it should be restricted or prohibited, depending on the needs of your particular audience.

Billing

Billing takes into account either the duration of the analyzed video or the length of time up to the stop tag, if the condition was triggered during the analysis.

What objects can be detected?

Gcore's AI nudity moderation can identify explicit content, involving either real individuals or deepfakes, as well as anthropomorphic cartoon characters. It is capable of detecting the following objects, whether they are fully or partially visible:

ANUS_EXPOSED

BUTTOCKS_EXPOSED

FEMALE_BREAST_EXPOSED

MALE_BREAST_EXPOSED

FEMALE_GENITALIA_EXPOSED

MALE_GENITALIA_EXPOSED

Analyze videos for nudity

Gcore's AI nudity detection is compatible with any MP4 video. This means you can upload files to Gcore Video Hosting or use MP4 videos from external storage sources. Nudity detection is not limited to videos; you can also detect explicit content in JPEG format images. For billing purposes, a picture will be treated as a one-second video.

Step 1. Obtain an MP4 link

If you're using a third-party video storage service other than Gcore, obtain a link to the MP4 video. If you're using Gcore's Video Hosting, execute the following API request to obtain the MP4 link:

      Copy
      
    curl -L 'https://api.gcore.com/streaming/videos/{video_id}' \
-H 'Content-Type: application/json' \
-H 'Authorization: APIKey {your_api_key}' \

Replace the example values {video_id} and {your_api_key} with the MP4 video URL and your actual API token.

Copy the value of the mp4_url field from the response.

      Copy
      
    curl -L 'https://api.gcore.com/streaming/videos/4428031' \
-H 'Content-Type: application/json' \
-H 'Authorization: APIKey {your_api_key}' \

      Copy
      
    {
    "id": 4428031,
    "name": "Sprite Fright, Blender (cut 30 seconds version for AI subtitles generation demo in API docs)",
    ...
    "hls_url": "https://demo-public.gvideo.io/videos/2675_G8EuYJFTsvIyWmKt/master.m3u8",
    "iframe_url": "https://player.gvideo.co/videos/2675_G8EuYJFTsvIyW",
    "iframe_embed_code": "<iframe width="560" height="315" src="https://player.gvideo.co/videos/2675_G8EuYJFTsvIyWmKt" frameborder="0" allow="autoplay; encrypted-media" allowfullscreen></iframe>",
    "converted_videos": [
        {
            "name": "vod720n",
            "width": 1720,
            "height": 720,
            "mp4_url": "https://demo-public.gvideo.io/videos/2675_G8EuYJFTsvIyWmKt/720.mp4"
        },
        {
            "name": "vod480n",
            "width": 1146,
            "height": 480,
            "mp4_url": "https://demo-public.gvideo.io/videos/2675_G8EuYJFTsvIyWmKt.mp4"
        },
       ...
    ]
}

Step 2. Create a nudity detection task

Execute the following API request to create a task:

      Copy
      
    curl -L 'https://api.gcore.com/streaming/ai/detect-nudity' \
-H 'Content-Type: application/json' \
-H 'Authorization: APIKey {your_api_key}' \
-d '{
    "url": "https://demo-files.gvideo.io/{your_video_url.mp4}" 
}'

Replace the example values {your_api_key} and {your_video_url.mp4} with the MP4 video URL and your actual API token.

In the response, you will get a unique task ID. It'll come in handy in the next step. For example: "task_id": "abc12345-6def…"

      Copy
      
    curl -L 'https://api.gcore.com/streaming/ai/detect-nudity' \
-H 'Content-Type: application/json' \
-H 'Authorization: APIKey 1234$abcd...' \
-d '{
    "url": "https://demo-files.gvideo.io/apidocs/spritefright-blender-cut30sec.mp4" 
}'

      Copy
      
    {
    "task_id": "34d0ba5a-23f6-4462-85ff-fe343b0cdb2d"
}

Pause detection if a stop tag is triggered

An object detected at the beginning of the video may immediately indicate that further analysis of the video is unnecessary. For such instances, you can add stop tags in the body of the nudity detection creation task.

Use the stop_objects parameter to specify stop tags, separating each one with a comma. You can also set a probability threshold value, expressed as a percentage. If this value is exceeded, the stop tag will be triggered.

Response sample:

      Copy
      
    curl -L 'https://api.gcore.com/streaming/ai/detect-nudity' \
-H 'Content-Type: application/json' \
-H 'Authorization: APIKey {your_api_key}' \
-d '{
    "url": "https://demo-files.gvideo.io/{your_video_url.mp4}",
    "stop_objects": "MALE_GENITALIA_EXPOSED:0.8,FEMALE_GENITALIA_EXPOSED" 
}'

Ensure you replace the example values {your_api_key} and {your_video_url.mp4} with the MP4 video URL and your actual API token.

Step 3. Get the task result

Wait a few minutes for the AI to complete nudity detection. To get the task result, execute the following API request:

      Copy
      
    curl -L 'https://api.gcore.com/streaming/ai/results/{your_task_id}' \
-H 'Content-Type: application/json' \
-H 'Authorization: APIKey {your_api_key}' \

There are five potential statuses (PENDING, STARTED, SUCCESS, FAILURE, and REVOKE) that reflect the progression of AI video processing. You should wait for the final statuses: SUCCESS or FAILURE.

As a result of the processing, you will receive a JSON file containing an array of entries if explicit content is detected. This final text result can be used to configure the logic of processes (for example, deciding whether to publish a video or not) on your backend. Below, we've provided an example of how this could be implemented.

      Copy
      
    {
    "task_data": {
        "url": "https://12345.gvideo.io/videos/12345_ZGNUY4U2NkhlF4xW/qid87v1_1080.mp4"
    },
    "task_id": "34d0ba5a-23f6-4462-85ff-fe343b0cdb2d",
    "processing_time": {
        "completed_at": "2024-03-26T07:00:07Z",
        "started_at": "2024-03-26T07:00:03Z",
        "total_time_sec": 3
    },
    "progress": 100,
    "status": "SUCCESS",
    "result": {
        "detection_results": [
            {
                "class": "FEMALE_BREAST_EXPOSED",
                "score": 0.5,
                "frame-no": 5
            }, …
         ]
    }
}

If there’s no nudity detected the array will be empty:

      Copy
      
    {
      "detection_results": []
}

The following response elements may be useful for further video moderation:

• class—a nudity type; see more details in What objects can be detected, for example, FEMALE_BREAST_EXPOSED.
• score—the probability that the object does relate to nudity, for example, 0.5.
• frame-no— the video frame number where nudity was found, for example, 5.

Assessing the probability of inappropriate content

There's no one-size-fits-all criterion or nudity score that can definitively determine whether a video is inappropriate. Different video hosting services cater to specific audiences such as adults, children, educational groups, etc. For instance, an acceptable nudity percentage for a site dedicated to sex education would be higher than for a hosting site that uploads entertainment videos intended for children.

You can set a probability threshold to determine when a video is inappropriate for your specific use case. One method is to run videos for one day and analyze the resulting probability coefficient.

How to use results in practice?

Here's a procedure for utilizing Gcore's AI nudity detection when checking user-generated content (UGC):

1. A user uploads a video to your app. You can use a platform like Gcore Video Hosting for video storage, but other platforms are also acceptable.

2. During the upload process, the video is transcoded and simultaneously checked for nudity (if you're using Gcore Video Hosting). If the video is from external storage, simply send us a link for AI nudity moderation.

3. The backend logic for deciding whether to publish is based on two parameters:

• The video has been transcoded and is now viewable.
• The video has been scanned for nudity (based on the analyzed coefficient) and no explicit content has been detected.