Research on an Improved YOLOv8 Object Detection Algorithm Based on MobileNetv4

The lightweight convolutional neural network designed for mobile devices features fast inference speed but is constrained by its inherent locality. Local information can only be captured within a windowed region, leading to performance degradation. Introducing the self-attention mechanism can capture global information, but it reduces detection speed. To address these issues, this paper introduces a hardware-friendly MobileNetv4 network architecture based on YOLOv8, incorporating a universally inverted bottleneck search block that integrates the inverted bottleneck, ConvNext, Feed Forward network, and a novel variant of extra depthwise convolution. Additionally, a dynamic upsampling operator is introduced to improve the upsampling operation, reducing GPU memory usage and latency in the model. Furthermore, this paper enhances the detection head of YOLOv8 by introducing a dynamic detection head, which combines spatial awareness, scale awareness, and task awareness into a unified framework. It effectively applies the attention mechanism in the object detection head, improving detection performance and efficiency. The experimental results demonstrate that compared to the next-best model, YOLOv8n, YOLOv8n_M achieved an improvement of 1.3% in mean Average Precision (mAP 0.5∶0.95). In terms of model complexity, YOLOv8n_M successfully compresses the model size by 36% (with a reduction of 1 million parameters) and reduces computational costs by 26% (The Giga Floating-Point Operations (GFLOPs) were reduced by 2.4.). The proposed YOLOv8_M effectively reduces the model's parameter count and inference time while improving object detection accuracy in various environments to a certain extent.