YOLOv8-C2fSE- An Improved Accuracy Model for Object Detection

The advancements in deep learning, and one of its important applications, are surveillance applications, where a lot of emphasis is placed on the object detection task, where the accuracy parameter is the most important factor, which always remains challenging. In terms of object detection tasks, we will think of YOLO, which is the most optimized technology for various tasks, including object detection. Since YOLOv8 is more modular and adaptable than other iterations, it is simpler to modify and adjust to particular use cases. YOLOv8 successfully strikes a compromise between speed and accuracy by utilizing new computing blocks, such as the C2f block, and sophisticated backbone networks. This preserves real-time processing capabilities while improving mean Average Precision (mAP) ratings across several datasets. The custom module, C2fSE (Squeeze-and-Excitation Module Embedded into a Cross-Stage Partial Block with the two convolutions (C2f)), is an improvised version of C2f, and developing the model called YOLOv8-C2fSE, which will help in achieving the channels-wise attention mechanism by prioritizing the important features, here we are deploying the module in the backbone and neck of the YOLOv8 architecture. One more additional change is to change the type of upsampling type to bilinear in the architecture of YOLOv8 for better gradient propagation and improved object localization. This YOLOv8-C2fSE architecture is tested on the MS-COCO data set, a benchmark dataset in computer vision technology. The experimental results show that our YOLOv8-C2fSE gives a better mAP50 percent of $\mathbf{5 3. 6}$ and an mAP50-95 percent of 38.1.