微型电脑应用

目标检测指对图像内的物体类型进行识别并且定位。一阶段目标检测算法从深层网络输出的特征图中获得分类信息与目标位置信息，然而深层特征由于经过长距离卷积与下采样处理而缺乏空间信息。针对该问题，参考语义分割的思想，通过将骨干网络中的浅层特征与上采样得到的深层特征结合的方法，对一阶段目标检测YOLOv5算法进行改进，利用骨干网络ResNet-50可以对特征图信息进行有效提取。在残差块中引入注意机制，对浅层特征提取阶段进行有效选择对象信息，并且可以将更多的权重分配给小的和弱的对象，以改善特征表达准确探测小物体的能力。另外，根据行人检测的数据特点，对预选框的生成方式和损失函数进行改进。采用INRIA和Caltech这2个数据集进行实验，结果表明，提出的改进模型在检测效果与检索速度方面均有所提升。

Target detection refers to the identification and positioning of object types in the image. The one-stage target detection algorithm obtains classification information and target location information from the feature map output by the deep network. However, the deep features lack spatial information due to long distance convolution and down-sampling processing. To solve this problem, refer to the idea of semantic segmentation. The YOLOv5 algorithm for target detection in the first stage is improved by combining the shallow features in the backbone network with the deep features obtained from up sampling. The backbone network is Resnet-50 to effectively extract the feature map information. Attention mechanism is introduced into the residual block to effectively select the object information in the shallow feature extraction stage, and more weights can be allocated to small and weak objects to improve the ability of feature representation to detect small objects accurately. In addition, according to the characteristics of pedestrian detection data, the generation method and loss function of the preselected box are improved. The experimental results on INRIA and Caltech data sets show that the proposed improved model improves the detection effect and retrieval speed.