Overview / Usage

Workplace sexual harassment, a pervasive societal issue, challenges individuals and organizations alike. In recent decades, it has emerged as a critical concern jeopardizing employee safety and well-being.

Ranging from subtle gender-based discrimination to overt assault, this issue emotionally and physically distress victims, harming productivity and satisfaction. Effective detection is crucial.

Various efforts, including legislation, education, and campaigns, have fallen short in preventing workplace harassment. It remains prevalent despite attempts to create safer environments, particularly affecting female employees with reported cases.

This underscores the need for innovative technological solutions that can identify and curb sexual harassment in real time, filling the gap left by traditional methods.

Methodology / Approach

Our solution aims to address the issue of sexual harassment in the workplace by utilizing deep learning techniques, particularly Convolutional Neural Networks (CNNs).

The solution's technical side entails video preprocessing, CNN architecture construction, training with loss functions, and metric-based evaluation. Transfer learning could enhance outcomes.

The iterative process involves experimentation, optimizing performance, refining architecture, and utilizing pre-trained models for accurate workplace sexual harassment detection.

We will follow an 80:20 split, where 80% of the images will be used for training our model, and the remaining 20% will be reserved for testing the model.

To further enhance the accuracy of the detection system, we have implemented Transfer Learning techniques by leveraging pre-trained models such as VGG16, Xception, and others.

These models have been trained on large-scale datasets and come with high accuracy levels.

Data preprocessing converts videos to frames and resizes them. CNN layers form the architecture, detecting patterns like edges, textures, and shapes via spatial convolution—separable convolutions lower parameters. Batch normalization normalizes outputs, reducing covariate shifts. These layers combine features, utilizing a sigmoid activation function for binary classification (harassment or non-harassment).

The model is trained using input frames and backpropagation to update weights and minimize loss. Validation data tracks performance during training. The final model is evaluated on unseen test data using accuracy, precision, recall, and F1-score metrics. Pre-trained models like VGG16, DenseNet201, and Xception are explored, using transfer learning to improve performance.

The solution's effectiveness is measured in terms of accuracy, precision, recall and F1-score, using evaluation metrics.

Technologies Used

•OneAPI

•OneDNN

•Intel(R) Extension for Scikit-learn

•Open CV

•Convolution neural network (CNN)

•Tensorflow

•Pandas

•Keras

•Jupyter Notebook

Documents and Presentations

Repository

https://github.com/Altaf-01/Sexual-Harassment-Detection-In-Workplace

Other links

• Harassment Detection Web Application

Collaborators