Preliminary summary of significance detection

At present, the relative traditional significance detection methods are mainly divided into two categories:

A class of data-driven, task-independent, bottom-up detection methods (mainly based on biological principles), and one that is dominated by consciousness-dependent tasks (mainly based on machine learning).

A method of image significance detection from top to bottom

The top-down significance model generally includes two parts: feature learning and significance calculation.

Due to the human top-down choice of attention by the high-level brain information control, for the same scene different people pay attention to different results, its motives, emotions and other factors more difficult to control and analysis, so the construction of the top-down of the significant model is more complex, so the study of such a significant model is not a lot.

A method of image significance detection from bottom to top

1. Feature extraction: Commonly used color characteristics, texture characteristics, shape characteristics, motion characteristics, local characteristics, image frequency (relatively new), etc.

Color characteristics: The basic color of the digital image is RGB, the other colors commonly used to indicate that the space has cmyk,cie Xyz,cie luv,cielab,yuv,hsv,his and so on.

Cielab color feature is the most commonly used image feature in the detection of significance, and it is also the most efficient feature in the detection of significance so far. , because the Cielab color model separates the luminance and chroma components closer to human visual characteristics. Studies have shown that visual sensitivity to brightness is far less than chroma, so it can be inferred that the sensing neurons of chroma and brightness are independent of each other. This is consistent with the Cielab color model.

2. A feature map is obtained by selecting several of the above features.

3. After that, the characteristics of the significant calculation, including multi-scale comparison method, regional comparison method, the global mean comparison method and so on.

Significance test criteria

1) The most prominent object

2) Consistent highlighting of the entire significant object

3) precisely match the boundary of the object

4) High noise resistance

5) Full resolution

An example of the significance detection method at the bottom can take a look at the classic method of Itti

Application

A brief look at some of the papers on the significance of the application of detection, there are the following application examples:

image Compression : The area of interest in the image is obtained by means of the significant detection method, and the area of interest in the compressed image is higher than that of the non-interest region, and the compression quality is improved by using different compression methods.

Video Significance detection : Consider the correlation between adjacent frames, and calculate the degree of significance of each region through intra-frame calculations and comparisons between adjacent frames.

The significance of video detection can also be used in video target tracking, motion detection, video compression and other fields.

Auto-keying diagram : A significant calculation of the image, roughly segmentation of the foreground and background, so as to automatically generate the label foreground, background, unknown area of the three-color map; Finally, the key graph algorithm can be used without manual automatic keying.

Biological principles related to the significance of detection

Also attached are some important biological principles of significance testing

Center-surround principle

A typical visual neuron is sensitive to only one small area, which is called the center. If stimulation is also generated in the area around the center, the stimulation inhibits the center's stimulation of the visual neurons. This structure makes the visual neurons more sensitive to the local space discontinuity, and is especially suitable for detecting the position which is more prominent than the local surroundings. This is how the retina, the lateral geniculate body and the primary visual cortex work. This can be said to be the theoretical basis for significant detection using local contrast.

Two-color opposing principle

There is a central area in the cerebral cortex that is more receptive to signals, one that excites the two colors while stimulating the region, and another that suppresses its excitement. This mechanism is called a two-color confrontation. In the area around the neurons, the opposite is the case. This shading mechanism, which exists between green and red, yellow, and blue, is usually established on the basis of the green and red, yellow, and blue color features of the broadband tuning.

The principle of contrast

Contrast can be on a texture or shape or color, the size of the contrast reflects the extent of the difference between an area and its surrounding areas, so the larger the contrast of the object, the greater the significance between them, the more can attract people's attention.

Preliminary summary of significance detection