Military Surveillance with Deep Convolutional Neural Network

摘要

In this paper, we propose a real time object detection system for images in Military Surveillance area. In traditional methods of object detection, region proposals are generated followed by extraction of features. At the end a classifier runs on these proposals. The speed of the process is slow and the accuracy is unsatisfactory. We propose a new model which provide better and higher mean average precision (mAP). The customization of YOLO model has enhanced Convolutional Neural Network (CNN) layers to 58. The customized dataset contains 22 classes which include 20 classes of Pascal VOC and 2 classes of tanks and guns from internet source. Challenging images like night vision, low resolution, images captured in different climatic conditions like fog, rain and snow are captured to test the model. The proposed model provides 79.12 % and 78.19% mean average precision (mAP) as compare to YOLOv2 model which provides 75.82% and 74.23% mAP using Pascal VOC and customized dataset respectively with input images of size 416*416. The customized Model is validated with different input size images for finding mean average precision with own dataset. We have customized YOLOv2 model by enhancing CNN layers and hyperfine tuned YOLOv2 model. We have used multi-GPUs for faster model training. We achieved ~33 times and ~50 times speedup as compare to CPU (Xeon) using single GPU (Tesla K40) and two GPUs architecture respectively.