实时语义图像分割模型研究

Research On Real-time Semantic Image Segmentation Model

为了实现快速语义图像分割,提出一种简化整合模型.首先,对频域视觉注意模型PQFT的四元数图像虚部系数进行简化改进.然后,将改进PQFT模型的显著图与简化PCNN的内部活动项结合起来对显著目标区域进行粗略定位,并以提出的显著目标区域均值的3/2倍进行精细分割.最后,根据"尺寸变化与否"准则判断输出正确的语义图像分割结果.实验结果表明,提出的整合模型具有实时性,且取得的AUC值和F值较原PQFT模型分别提高了29.9%和44.2%.

Aiming at addressing fast semantic image segmentation,a simple integrated model was proposed.Firstly,the PQFT model,a frequency-domain visual attention model,was improved by improving imaginary coefficients of its quaternion image.Then,the saliency map of the improved PQFT model was integrated with the inner activity of a simplified PCNN to locate the raw salient region,and the detected salient object was segmented perfectly according to the proposed 3/2times meanvalue threshold method.At last,the accurate semantic image segmentation result was output according to the size-changing rule.The experimental results show that the fast semantic image segmentation model proposed is of real-time,and its＇ AUC and F values have been increased 29.9% and 44.2%,respectively.