基于分部特征计算的轻量化非结构目标检测

Efficient object detection method with partial calculation for unstructured scenes

针对非结构化场景(工地、矿场)缺少特殊目标的通用数据集、复杂特征难以准确提取以及计算复杂度高的问题,构建了一个面向非结构场景的特殊目标检测数据集,并提一种轻量化目标检测模型YOLO-PT,以极低的计算量达到了较高的检测精度。通过构建分部特征计算(partial feature calculation, PFC)模型减少特征冗余信息的计算,并引入了多头自注意力机制来增强复杂特征的提取精度,同时设计多通道金字塔结构对多尺度特征进行渐进式融合,提高复杂对象的识别精度。最后在非结构场景进行实验验证,结果表明,所提出方法仅在4.3×10^(6)的参数量下就达到了53%的准确率,在精度、参数量以及浮点运算量上均优于其他方法。

To address the challenges of the absence of shared datasets covering unique targets in unstructured scenes(such as construction sites and mining sites),the difficulty in precise extraction of complex features,and the high computational complexity,this paper creates a dedicated object detection dataset for unstructured scenes.We present a lightweight object detection model named YOLO-PT,which attains high detection accuracy while requiring minimal computational resources.We mitigate the computation of redundant feature information by developing a partial feature calculation(PFC)model.We also incorporate a multi-head self-attention mechanism to enhance the precision of complex feature extraction and design a multi-channel pyramid structure for the gradual fusion of multi-scale features,thereby improving the recognition accuracy of complex objects.Finally,experimental validation is conducted in unstructured scenarios.The results demonstrate that the method proposed achieves the accuracy of 53%with a mere 4.3×10^(6) parameters,outperforming other methods in terms of accuracy,the number of parameters and floating-point operations.