Development of super resolution based hybrid model for the detection of dangerous objects

The presence of dangerous objects carried or used by unauthorized persons in public spaces is known to threaten public safety. Authorities have camera systems such as CCTV and MOBESE in public spaces to ensure security. However, since these systems are monitored 24/7 by operators, possible human errors may occur. For this reason, it is very important to develop systems that help operators to automatically recognize dangerous objects. Accordingly, in this thesis, a hybrid model consisting of YOLO, which detects the targeted object on digital images and videos, and SR algorithm, which enhances the image, is proposed to detect dangerous objects. For the training of the proposed hybrid model, a new dataset called DMGDATA consisting of 18,231 data was prepared. With this dataset, selected versions of the YOLO architecture were trained and according to the metrics obtained, it was determined that the YOLOv8m model was suitable for the thesis. For the SR layer in the proposed hybrid model, a new algorithm called DCASR, which consists of channel attention mechanism and residual block layers, is proposed. When this algorithm is compared with its similar algorithms, it is observed that it produces more successful results in terms of speed and PSNR value. The images enhanced with the DCASR model were sent to the YOLOv8m model for detection after removing the noise with the BF filter. The proposed YOLOv8m+DCASR+BF model was tested with a test set of two different resolutions and different SR algorithms. As a result of the test procedures, the hybrid model created using the DCASR algorithm was found to be more successful than the others. With the proposed hybrid model, object detection performance in normal resolution images is not improved, while object detection performance in low resolution images is improved by 4% to 6%. In this way, it is thought that images from low-quality sources can perform more successful object detection using algorithmic methods.