A novel model for higher performance object detection with deep channel attention super resolution

Abstract

With the introduction of deep learning methods, target object detection studies have gained momentum and started to be used in many areas. In recent years, various problems related to training the model with appropriate images and improving the current image quality to detect the target object more successfully have been discussed. In the light of these discussions, the identification of dangerous objects, which is one of the most important areas of object detection, is quite remarkable. For this reason, within the scope of study, a new dataset called DMGDATA-mini was prepared to detect dangerous objects and presented to the public. By using the data in the training of the You Only Look Once (YOLO) model of YOLOv5, it was ensured that the model successfully detected the object. In order to increase the object detection success and observe the results, popular Super Resolution (SR) algorithms and the Deep Channel Attention Super Resolution (DCASR) algorithm developed by us were integrated into the structure and the effect of SR algorithms on object detection was observed. It was determined with the Peak Signal-to-Noise Ratio (PSNR) metric that the developed DCASR algorithm performs better image enhancement than all SR algorithms in the study. Also, YOLOv5 and YOLOv5 + SR models were also compared. Thanks to the improved DCASR model, a 9.9 % increase in object detection success was observed. It was observed that SR algorithms have a positive effect on performance in general.