针对序列图像内具有低信噪比和低对比度特征的运动目标, 提出了一种基于级联分类器的弱小目标检测算法。该算法从安瓿瓶序列图像内提取绝对差分值、局部差分对比度和局部相关系数3个图像特征。每个图像特征对应一个分类器, 通过三层级联形式实现序列图像中的小目标检测。第一个节点与传统帧间差分法类似, 主要去除大量背景图像并检测出大颗粒运动目标, 后两个节点则用于检测弱小目标、排除光流和瓶身污渍产生的噪声点。实验结果显示, 相对于传统的帧间差分法, 本文算法具有高检测精度和高抗干扰能力等特点, 不仅可以检测出图像中弱小运动目标, 同时也消除了复杂背景下的噪声影响, 弱小目标的检出率达到99.3%, 并且满足安瓿在线检测的实时性要求。

An inspection algorithm based on cascade classifiers is presented for detecting small moving foreign substances with low Signal and Noise Ratio (SNR) and low contrast in sequential images. The algorithm obtains three features of absolute difference, local difference contrast and neighborhood correlation from the sequential images of an ampoule. Each feature corresponds to a classifier, and small foreign substances are inspected by using three-layer cascade classifiers. The first layer corresponds to a traditional frame differencing method, which is used to remove the background and detect the large moving foreign substances. The next two layers are used to inspect small foreign substances and remove the noises generated by optical flow and the stain of bottle. Experiment results show that compared with the traditional frame differencing method, this algorithm has higher detection precision and higher anti-interference ability in inspecting small substances with the interference of a complex background, and the detection rate of small foreign substance is 99.3%. This algorithm can meet the requirement of real-time detection of ampoules for medicine production.