基于注意力机制和特征融合改进的小目标检测算法

AN IMPROVED SMALL OBJECT DETECTION ALGORITHM BASED ON ATTENTION MECHANISM AND FEATURE FUSION

近年来目标检测在机器视觉领域得到了极大的发展。然而对于图像中的微小物体,其所占像素少,容易受到背景因素等影响。当前算法在进行卷积采样时容易丢失小目标的特征信息,对于小目标达不到很好的检测效果。针对图像中小目标检测存在的问题,在研究当前目标检测算法的特点时发现SSD系列算法兼顾检测精度和速度的优点。在SSD网络框架中引入注意力模块,有效提取小目标的特征信息。使用特征融合的方式对小目标进行精确的位置回归。通过实验在TILDA织物瑕疵数据集和VEDAI航拍数据集上验证了该方法的可行性,对于图像中的小目标可以有效检测,同时减少了错检。

In recent years,target detection has been greatly developed in the field of machine vision.However,for tiny objects in the image,it takes up fewer pixels and is susceptible to background factors.The current algorithm is prone to lose the feature information of small targets in convolution sampling,so it can t achieve a good detection effect for small objects.Aiming at the problems of small target detection in images,this paper studies the characteristics of the current target detection algorithm and finds that SSD series algorithm has the advantages of both detection accuracy and speed.The attention module was introduced to effectively extract the feature information of small targets in the SSD network framework.The feature fusion was used to accurately locate the small target.The feasibility of this method is verified by experiments on TILDA fabric defect data set and VEDAI aerial photograph data set.It can detect small targets effectively in the image and reduce false detection.