Nudenet Onnx

The implementation includes two key components: a YOLOv8 detection model (320n.onnx) for generating initial bounding box proposals and a non-maximum suppression module (nms-yolov8.onnx) for post-processing and filtering duplicate detections. The system can identify 18 different anatomical regions and clothing states, including both exposed and covered body parts with high precision.

Performance is optimized through model quantization and ONNX Runtime integration, providing significant speed improvements over the original implementation while maintaining detection accuracy. The models operate on 320x320 input resolution and support configurable detection thresholds for balancing precision and recall according to specific application requirements.

Usage

Installation

pip install onnxruntime>=1.18.0
# For GPU acceleration (optional):
# pip install onnxruntime-gpu>=1.18.0
pip install dghs-imgutils

Basic Detection

from imgutils.detect.nudenet import detect_with_nudenet
from PIL import Image

# Load and process image
image = Image.open('your_image.jpg')
detections = detect_with_nudenet(
    image, 
    topk=100,
    iou_threshold=0.45,
    score_threshold=0.25
)

# Process results
for bbox, label, confidence in detections:
    x1, y1, x2, y2 = bbox
    print(f"Detected {label} with confidence {confidence:.3f} at [{x1:.1f}, {y1:.1f}, {x2:.1f}, {y2:.1f}]")

Original Content

ONNX models of notAI-tech/NudeNet.

Label Definitions

Here is a detailed list of labels from the NudeNet detection model and their respective meanings:

Citation

@misc{nudenet_onnx,
  title        = {{NudeNet ONNX Models}},
  author       = {deepghs and notAI-tech},
  howpublished = {\url{https://huggingface.co/deepghs/nudenet_onnx}},
  year         = {2023},
  note         = {Optimized ONNX models for the NudeNet object detection system, providing efficient nudity detection with YOLOv8 architecture},
  abstract     = {This repository provides optimized ONNX models for the NudeNet object detection system, originally developed by notAI-tech. The models are designed for efficient nudity detection in images using a YOLOv8-based architecture with specialized training for NSFW content recognition. These ONNX-format models enable cross-platform compatibility and accelerated inference across various hardware backends. The implementation includes two key components: a YOLOv8 detection model for generating initial bounding box proposals and a non-maximum suppression module for post-processing and filtering duplicate detections.},
  keywords     = {object-detection, onnx, yolov8, nsfw-detection}
}