.png){kind=icon}
.png){kind=icon}
Return
Rate-accuracy optimized quantization algorithm based on ROI image coding in power line inspection
.png){kind=icon}
PRE
AI
DOI：10.1007/s11042-023-15271-7.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In the context of big data, transmission line inspection of the Grid has evolved from the era of human inspection to intelligent inspection. Large quantities of image data will be collected and analyzed by machines. Since the foreground may have greater value compared with the background, Region of Interest (ROI) image coding is applied. However, the traditional image coding aims to maintain good human-perceivable visual quality and is not designed for semantic analysis. The image coding paradigms that comprehensively balance the human visual quality and automatic analysis performance are needed. In this paper, a Rate-Accuracy Optimized quantization algorithm based on Region of Interest image coding is proposed to obtain the optimal analysis performance with the given coded bit rate. First, a machine vision-oriented attention-map is determined. Since the features are leveraged to reflect abstract semantic meaning which is vital for image analysis tasks, it is reasonable to regard the region containing the feature vector information as the key area and others as the non-key area. The key area is compressed with fine-grained quantization while the non-key area is with coarse-grained quantization. Then the relationship between rate, accuracy and quantization parameters are analyzed and modeled. Finally, the optimal quantization parameters are determined based on the Rate-Accuracy criteria. The proposed algorithm is verified by the insulator dataset (Image of insulator defect collected by drone inspection). Experimental results show that the accuracy of defect identification is improved by 12% at the same bit rate.
Keywords:
Intelligent inspection
Image coding
Attention-map
Rate-accuracy
Journal
IF:
3
Papers:
1.9W
Citations:
3.2W
Organization
View more organizations