چكيده به لاتين
Detecting vehicles in images under adverse weather conditions, particularly during heavy rain, is one of the significant challenges in the fields of computer vision and intelligent transportation systems. This thesis aims to improve the accuracy of vehicle detection in rainy conditions by examining various methods and optimizations. Initially, the analysis introduces deep learning structures, especially the YOLOv9 model, as one of the latest architectures for real-time object detection. Key changes were then implemented in this model to enhance its performance. One of the most significant modifications made to the YOLOv9 structure was the addition of a three-layer attention block in the core section of the model. This attention block helps the model extract critical information and features from images with greater precision. As a result of this optimization, the model's detection accuracy increased from 62% to 69%, demonstrating the positive impact of these changes on model performance. Next, a dedicated dataset was created for this project, consisting of 3,500 images of vehicles in both rainy and non-rainy conditions, including both natural and synthetic images. This dataset was designed to improve modeling and ensure more accurate training of the network in rainy conditions. Finally, to enhance the quality of input images and improve detection accuracy, a Generative Adversarial Network (GAN) model was employed. This model increased the resolution and quality of images, leading to an improvement in the final accuracy of the proposed YOLOv9 model. Additionally, the Gray Wolf Optimizer (GWO) was utilized to optimize the model's hyperparameters. This algorithm, by simulating the social behavior of gray wolves, effectively narrowed the search space to find the optimal values for the parameters. The use of GWO improved the model's performance, ultimately achieving an overall vehicle detection accuracy of 83%. The results obtained from this research indicate that employing optimization techniques, such as adding attention blocks, utilizing a diverse dataset, and improving the quality of input images, can significantly enhance detection accuracy by up to 24%. These approaches can greatly contribute to improving the performance of intelligent transportation systems and other applications related to object detection.