An Event Detection Mechanism with Deep Feature Extraction and Optimal Loss Function Based XGBoost Classifier

Abstract

Purpose: Event Detection gains significant attention in retrieving information of NLP (Natural Language Processing) which could be applied in several areas. However, it becomes impractical to determine and classify the events based on features due to dynamic data nature, imbalanced data issue and high volume in image streams. To lessen these pitfalls, the study propounded to determine events from multiple image classes

Materials and Methods: The study proposes VGG-16 (Visual Geometry Group-16) and ResNet-50 (Residual Network-50) for feature extraction and optimal loss function based XGBoost (eXtreme Gradient Boosting) for classification. The VGG-16 model assess neural network and extracts features by increasing the architecture depth with smaller convolutional-filters, bringing out better improvement in feature extraction. However, some of the features when moving to deep layers find difficulty in propagating gradient information from inner layers from output back to input layers. The features from two neural network models are incorporated by feature fusion process using deep stacked auto-encoder method, consisting of multiple sparse autoencoder layers, wherein each layer’s input is connected to input of following network layer. Further, to address the overfitting issue, imbalanced dataset problem, classification error, and to focus on data loss, Optimal Loss Function with XGBoost classifier is used to perform effective event detection based on the training process. The detection of event that corresponds to event classes are performed through classified image features.

Results: The performance assessment of proposed model, enumerated through assessing event detection accuracy of 97.58% in comparison with other conventional methods.

Conclusion: The results explicated the superior performance of the proposed model, than other existing event detection model.