改进Yolov5s的移动端AR目标识别算法

Mobile AR target recognition algorithm based on improved Yolov5s

针对目标识别模型存在参数量大、识别速度慢的问题,提出了一种改进的轻量化目标检测算法Yolov5s-MCB。将MobileNetV3网络作为Yolov5s主干特征提取网络以降低模型参数量。为了更好地拟合非线性数据优化模型收敛效果,将MobileNetV3网络ReLU激活函数替换成Mish激活函数以避免梯度消失和梯度爆炸。增加BiFPN特征金字塔结构,利用一种迭代式的特征融合方法提高检测精度。此外,引入坐标注意力机制使得模型关注大范围的位置信息以提高检测性能。为了优化模型训练收敛速度,采用Focal-Loss EIoU作为边框回归损失函数来解决低质量样本产生损失值剧烈震荡的问题。实验结果表明,该算法在VOC数据集的平均识别精度达到了90.5%,模型大小为7.63 MB,检测速度为99 FPS,与原Yolov5s相比,在保持识别精度不变的情况下,推理速度提升了17.85%,模型大小降低了45.9%,满足检测任务的实时性和检测精度要求。同时,将Yolov5s-MCB模型转为ONNX模型移植到手机上,结合ARCore SDK开发一个附带目标检测功能的AR应用。

To address the problems of many parameters and slow recognition speed of existing target recognition models,an improved lightweight target detection algorithm Yolov5s-MCB is proposed.Firstly,MobileNetV3 network is used as the Yolov5s backbone feature extraction network to reduce the number of parameters of the model.In order to fit the nonlinear data better and optimize the model convergence effecter,the MobileNetV3 network frontal ReLU activation function is replaced by Mish activation function to avoid gradient disappearance and gradient explosion.Secondly,the BiFPN feature pyramid structure is added to improve the detection accuracy with an iterative feature fusion method.In addition,the introduction of coordinate attention mechanism allows the model to focus on a wide range of location information to improve the detection performance.In order to optimize the model training rate of convergence,Focal-Loss EIOU is used as the border regression loss function to solve the problem of low-quality samples generating drastic oscillations in loss values.The experimental results show that the algorithm achieves an average recognition accuracy of 90.5%in the VOC dataset,a model size of 7.63 MB,and a detection speed of 99 FPS.Compared with Yolov5s,the proposed algorithm improves the inference speed by 17.85%and reduces the model size by 45.9%while keeping the recognition accuracy unchanged,meeting the requirements of the real-time detection tasks and detection accuracy.And the Yolov5s-MCB model is converted to ONNX model and ported to a cell phone to develop an AR application with target detection function in combination with ARCore SDK.