Improved real-time object detection method based on YOLOv8: a refined approach

With the rise of deep learning, researchers have proposed numerous advanced object detection algorithms. However, these often face challenges such as high parameter count and complexity. While some existing methods have reduced model parameters and complexity, they may sacrifice accuracy, reducing practical value. Therefore, achieving an optimal balance between maintaining detection accuracy and real-time performance has become a paramount concern for researchers. This paper introduces an enhanced YOLOv8 object detection model to tackle this issue. We propose the Faster-C2f module, inspired by RepViT, and the Rep-Fasterblock network based on FasterNet. By reparameterizing the FasterNetBlock module with Partial convolution and Depthwise separable convolution, our model effectively reduces parameters while efficiently handling spatial and channel information in the feature map. We enhance SPPF by combining AdaptiveMaxPool2d and AdaptiveAvgPool2d at different scales, introducing the SimAM attention mechanism to obtain Sim-SPPF. This improvement enables the network to capture features at various scales while increasing sensitivity towards key features. Additionally, we propose two neck network structures: enhanced PANet and CNN-based Cross-scale Feature PANet (CCFPANet). The PANet model is enhanced with our novel convolution PsConv and three CSPCx of varying complexity levels to reduce parameters while maintaining efficiency. Additionally, CCFPANet effectively combines detailed features with contextual information across scales. Furthermore, we introduce a decoupled detection head where Partial convolution replaces the traditional 3 \(\times\) $\times$ 3 Conv operation to further reduce parameters. Our proposed YOLOv8PANet and YOLOv8CCFPANet have different neck network structures and achieve a 21%/54.5% reduction in parameter count compared to YOLOv8 while maintaining detection accuracy of 84.2%/83.3% on the PASCAL VOC07+12 dataset and achieving detection accuracies of 42.1%/40.0% on the MSCOCO dataset respectively; furthermore, our experiments demonstrate high-speed detection trials reaching 100 FPS/ 108 FPS. These results indicate that our method achieves a superior balance between detection accuracy and real-time performance compared to existing methods, positioning it as a competitive solution in the field of real-time object detection.