Main color recognition method for target object in image

Abstract

The invention provides a main color recognition method for a target object in an image, which relates to the field of image recognition. The method comprises steps: firstly, the divided type number of RGB color space color values is determined, clustering is then carried out, and a color mapping table between a color value and a color type is generated; as for each image in a specified data set, the area of the target object in a selected image is positioned, the image of the area is converted to HSV space, clustering is then carried out and then the image is converted back to the RGB space; according to a counting rule, a different weight is given to a pixel point at each different position in the area of the target object, a weighted color histogram of the area is obtained, the color corresponding to the highest rectangle in the histogram is regarded as the recognized main color of the target object, the color type corresponding to the main color is searched in the color mapping table, and a marking result is recorded. The color recognition accuracy and the color marking extensibility are enhanced.

Description

technical field [0001] The invention relates to the field of image recognition, in particular to a main color recognition method of a target object in an image. Background technique [0002] Color is a main attribute of the target object itself, and it is one of the main factors for machines and humans to distinguish different target objects. Recognizing the color of the target object in the image can be used to more accurately distinguish objects of the same category, and can also be applied in the data preparation process in machine learning, greatly improving the speed of data labeling and saving the time and labor costs of manual labeling. According to whether the material can resist deformation, the target object can be divided into two types: rigid object and flexible object. For the target object, especially for the flexible object composed of multiple materials and elements, on the one hand, the target object or its background itself has rich colors, so that the tar...

AI Technical Summary

Problems solved by technology

[0007] (1) When the data scale is large, the manual labeling period is long and the cost is high;

[0008] (2) It is necessary to know the color distribution of the target object in the data set to be labeled in advance, and the requirements for the data itself and the preparation work before labeling are too high; statistical errors will greatly affect the quality of labeling;

[0009] (3) prone to labeling errors

[0010] (4) Labeling standards are different, and the results may be contradictory

[0015] (2) The annotation results of different batches of the same category cannot be simply combined

Since the main color category is related to the dataset, the color annotation results of different batches of datasets cannot be directly merged

If you want to merge data sets, you need to label all the data again, which brings a lot of inconvenience to the incremental expansion of data sets

[0016] (3) Octrees tend to cause uneven granularity of color categories

[0021] (2) The number, coverage and granularity of color categories are not easy to determine in advance

When manually determining the color category to be marked, there is a lot of subjectivity, and it is extremely prone to problems such as incomplete color distribution and uneven color granularity

[0022] In short, there are some unavoidable shortcomings in the existing main color recognition technology of complex target objects

The main color recognition problem of complex target objects is characterized by a huge color space but uncertain color distribution of data set images, complex color types of target objects in the image and uncertain main color and non-dominant color distribution of the area where the target object is located, and image illumination changes Variety, it is not easy to determine the label color category of the main color, etc.

Images