机器视觉中小目标检测实验优化模型设计与实现  

Design and Implementation of Optimized Model for Small Object Detection in Computer Vision Course

作　　者：孙楠 杨煜戎 杨哲[1] 卜子渝 SUN Nan;YANG Yurong;YANG Zhe;BU Ziyu(School of Computer Science and Technology,Jiangsu Provincial Key Laboratory for Computer Information Processing Technology,Soochow University,Suzhou 215006,Jiangsu,China)

机构地区：[1]苏州大学计算机科学与技术学院,江苏省计算机信息处理技术重点实验室,江苏苏州215006

出　　处：《实验室研究与探索》2023年第3期32-39,共8页Research and Exploration In Laboratory

基　　金：江苏省高等学校自然科学研究重大项目(19KJA230001);江苏省高等教育教改立项研究课题(2017JSJG001)。

摘　　要：目标检测是计算机视觉课程的重要实验内容之一,但现有模型对小目标检测能力普遍较弱。为了加深学生对现有模型结构和缺陷的深入理解,掌握模型的优化方法,基于实验中常用的SSD模型,引入了轻量级秩扩展网络ReXNet,重新设计了特征融合与过滤模块。特征融合模块在深浅层特征融合之前,先对深层特征图进行特征抽取,减少无效语义信息对浅层特征的干扰,增强了模型对小目标语义特征的表征能力。特征过滤模块则分别在分类和回归时,引入通道注意力和空间注意力的双路结构,提高分类与回归的精度。在VOC和COCO数据集上的实验结果表明,改进后的模型不仅提高了对小目标的检测性能,保留了较快的检测速度,而且改善了原始模型存在的漏检问题。通过模型设计的优化,加深了学生对于目标检测模型架构的理解,提高了学生的综合实践能力,促进了计算机视觉课程的实验教学内容建设。Object detection is one of the important experiments in computer vision course,but the existing models are generally weak in detecting small targets.In order to deepen students’understanding of the existing model structure and weaknesses,and master the optimization method,this paper replaces the backbone network of SSD model with the lightweight rank extended network ReXnet,and redesigns the feature fusion and filtering module.The feature fusion module extracts features from deep feature maps to reduce the interference of invalid semantic information on the shallow features,and enhance the representation ability of the model to the small objects.The feature filtering module introduces the two-way structure of channel attention and spatial attention to improve the accuracy of classification and regression.The experimental results on VOC and COCO data sets show that the improved model not only improves the detection performance of small targets,retains the faster detection speed,but also improves the missed detection problem of the original model.Through the optimization and thought of model design,students’understanding of the structure of target detection model is deepened,students’comprehensive practical ability is improved,and the experimental teaching construction of computer vision course is promoted.

关 键 词：目标检测 实验设计 特征融合 特征过滤 秩扩展网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]