基于改进YOLO_v3算法的车辆轮胎紧固件检测方法优化

Optimization of Vehicle Tire Fastener Detection Method Based on Improved YOLO_v3 Algorithm

轮胎拆装是汽修行业修补轮胎必要步骤之一,由于重型货车轮胎过重导致工人腰椎受到严重损伤。现将EfficientNet-B4轻量型网络结构算法替换YOLO_v3算法的主干部分(DarkNet-53),从而实现各类车型轮胎紧固件的识别。替换后的网络参数大量减少,紧固件样本训练速度加快。YOLO_v3使用二分类交叉熵损失函数对正负样本分类并计算损失,现使用Focal loss分类损失函数替换二分类交叉熵损失函数,从而提出一种新的神经网络模型(YOLO_v3-Nut)。实验结果表明,YOLO_v3-Nut模型在训练与识别速度特性上都更优于YOLO_v3模型,且文中的模型结构比YOLO_v3模型储存空间减少了43.01%,算法平均准确率(MAP)为93.2%,同时检测速度为36 fps,足够完成各类车型轮胎紧固件的识别。

Tire disassembly is one of the necessary steps for repairing tires in the auto repair industry.Due to the excessive heavy truck tires,the lumbar spine is damaged by the lumbar spine.The EfficientNet-B4 lightweight network structure algorithm is now replaced with the main part of the YOLO_v3 algorithm(DarkNet-53)to realize the identification of tire fasteners in various models.The replaced network parameters are largely reduced,and the fastener sample training speed is accelerated.YOLO_v3 uses a duplex cross-entropy loss function to classify and calculate the loss of positive and negative samples.The Focal loss classification loss function is now replaced by the binary cross-entropy loss function,so as to propose a new neural network model(YOLO_v3-Nut).The experimental results show that the YOLO_v3-Nut model is better than the YOLO_v3 model in terms of training and recognition speed characteristics,and the model structure of this article is reduced by 43.01%compared with the YOLO_v3 model storage space.The speed is 36 fps,which is enough to complete the identification of tire fasteners in various models.