Preprocessing Method Comparisons For VGG16 Fast-RCNN Pistol Detection

摘要

In recent years, gun detection and threat surveillance became a popular issue as gun violence continued to threaten public safety. Convolution Neural Networks (CNN) has achieved impressive gun detection precision with the advancements in graphic processing units. While many articles have proposed beneficial complex architectures within the neural network, there has been little study on effective image preprocessing techniques that supplement neural networks. With the objective of increasing neural net precision using image processing techniques, this research analyzes three different approaches to image preprocessing using a VGG16 trained Fast Regional Convolutional Neural Network (FRCNN) pistol detector. The base VGG16 was trained with transfer learning in MATLAB on a dataset of 1500 pistol images and tested on 500 more. The results of the original VGG16 detector are compared with the results of the other VGG16 detectors trained with various image processing techniques to determine the viability of each technique. The three image processing techniques are as follows, color contrast enhancement, principle component analysis (PCA), and a combined preprocessing method. After testing the detector trained with the three methods above, it was found that the color enhancement technique had the best success in raising precision with proper levels of color contrast adjustments. The PCA analysis proved to be incompatible for the neural net to learn features of images that has not underwent PCA processing and thus the method failed to produce beneficial results on the unmodified testing dataset. The combined method processing took both PCA and color contrast enhancement techniques and combined the results into a single training dataset. The combined preprocessing method proved to be ineffective in raising precision potentially due to conflicting features.