An Optimization of Multimodal Deep Crime Detection Network Using Osprey Optimization Algorithm

Sundari Palanisamy, Malathi Arunachalam

Abstract: Crime detection helps law enforcement to prevent future crimes by identifying their patterns. However, evolving criminal behaviors and rapid crime occurrences makes the future crime classification difficult. Various Deep Learning (DL) models are made to predict crime accurately. Among them, Social media information enriched Multimodal Diversified Deep Crime detection network (SMDDCnet) was proposed for crime prediction using historical dataset and social media (twitter) image and video information. This model uses enhanced recursive self-attention mechanism (ERSAM) to capture long term dependencies, Convex Function Information Entropy (CFIE) quantifies the uncertainty in crime patterns and Convolutional Bidirectional Long Short Term Memory (ConvBiLSTM) to forecast the crime effectively. But, the hyperparameters of ConvBiLSTM are not optimally selected causing lower accuracy result and higher computational complexity. Also, hyperparameter tuning is typically performed manually, making the process time-consuming and work-intensive. Hence, Social media information enriched Optimized Multimodal Diversified Deep Crime detection network (SMODDCnet) is proposed in this paper for tuning hyperparameters of the ConvBiLSTM network to enhance the crime prediction performance. The Osprey Optimization technique (OOA) is a novel metaheuristic optimization technique is employed in this study. It is based on how ospreys seek fish in the sea. The OOA involves two stages such as exploration and exploitation, wherein ospreys identify their prey's location and move it to an appropriate location to eat it. Based on these processes, the optimal hyperparameters of the ConvBiLSTM model are determined for the prediction of crime from both historical and social media information data. By utilizing the OOA, this model effectively fine-tunes the hyperparameter of ConvBiLSTM, resulting with high accuracy and lower computational complexity. Finally, the results of the experiment indicate that the SMODDCnet model is more accurate than other crime prediction models on the Crime in India dataset, with an accuracy rate of 98.42%.

Keywords: Deep Learning, Social Media, Crime Detection, Osprey Optimization