摘要
目标检测是机器视觉领域一个重要的基础性方向,是以标示出图像中感兴趣目标的真实位置为目的的工作,而图像中的部分目标往往处于被遮挡的状态。由于实际环境中目标被遮挡程度和遮挡目标物体性质不同等因素的影响,提高检测被遮挡目标的准确性是一个难点。该文以一步检测法SSD目标检测算法为基础,在部分卷积层结构中添加注意力机制模块CBAM,有目的地关注特征图中的重要信息以较好地检测被遮挡的小目标。还引入Inception-ResNet-v2网络结构改变SSD算法中特征图的生成方式,并嵌入CBAM改进Inception-ResNet-v2的部分结构,更好地提取关键信息以区分被遮挡目标和干扰项。分别从行人和车辆2个不同目标出发,混合开源数据集和自建数据集进行训练,测试结果表明改进后的模型检测被遮挡目标的效果有所提升。
Object detection is an important basic direction in the field of machine vision, which aims to indicate the real location of the objects of interest in the image, and some of the targets in the image are often occluded. Due to the influence of factors such as the degree of occlusion and the nature of occluded objects in the actual environment, it is difficult to improve the accuracy of detecting occluded targets. Based on the one-step detection SSD object detection algorithm, this paper adds the attention mechanism module CBAM to the partial convolution layer structure to pay attention to the important information in the feature graph in order to better detect the occluded small targets. In addition, the Inception-ResNet-v2 network structure is introduced to change the generation mode of the feature graph in the SSD algorithm, and CBAM is embedded to improve part of the structure of Inception-ResNet-v2 to better extract key information to distinguish between occluded objects and interference items. Starting with two different targets of pedestrians and vehicles, mixed open source data sets and self-built data sets are trained, and the test results show that the effect of the improved model in detecting occluded objects is improved.
出处
《科技创新与应用》
2023年第8期10-14,共5页
Technology Innovation and Application