Lightweight object detection network improved based on YOLOv8

In response to the scenarios where edge devices have the requirement for polarized image object detection, this paper presents a lightweight object detection network (P-faster-yolov8) improved based on YOLOv8. The FasterNet architecture is employed to decrease the FLOPs (floating point operations per second) and memory footprint. The CGLU (Convolutional Gated Linear Unit) module, which integrates convolutional operations and gated linear units, is utilized to enhance the efficiency of local information utilization. Specifically, a novel C2f_Faster_CGLU module is designed. This module not only realizes network light weighting but also enhances the effectiveness and flexibility of feature extraction. Furthermore, a Content-Guided Attention Fusion (CGAF) module is introduced to improve the network neck structure, thus optimizing the feature fusion process. Experimental results demonstrate that, in comparison with the original YOLOv8 network, the proposed network after lightweight improvement shows a 25.9% reduction in the number of parameters and a 25.7% decrease in the amount of computation.