Brief introduction of common features of content-based image retrieval system

1. Preface

Content-based image retrieval (contents Based image retrieval, CBIR) is a retrieval technique that finds images similar to retrieved content from an image database. It makes use of the image features, such as color, texture, contour and shape, which are automatically extracted from the image, which can be calculated and compared to retrieve the result image set according to the user's requirement, and the retrieval result may be improved by using the relevant feedback technology in the text retrieval technology. At present, the technology of image retrieval system is based on the calculation and comparison of the feature information of the underlying image, which is also called "visual similarity".

The core of Cbir is to retrieve images using the visual features of the images. In essence, it is a kind of approximate matching technology, which integrates computer vision, image processing, image understanding and database and other fields of technical achievements, in which feature extraction and indexing can be done automatically by computer, avoiding the subjectivity of manual description. The process of user retrieval is typically to provide a sample image (Queryby Example) or to depict a sketch (Queryby Sketch), the system extracts the characteristics of the query image, compares it with the features in the database, and returns an image similar to the query feature to the user. The implementation of Cbir relies on the resolution of two key technologies: Image feature extraction and matching. This paper mainly introduces three kinds of visual features commonly used in Cbir systems: color, texture and shape.

2. Color characteristics

Color is a visual feature of the surface of the object, each object has its own unique color characteristics, such as people often refer to the green trees or grasslands, refers to the blue often think of the sea or Blue sky, the same type of objects often have similar color characteristics, so we can according to color characteristics to distinguish objects. Generally use the histogram to describe the color characteristics, the histogram describes the different colors of the pixels in the whole image of the proportion, can also be understood as the color histogram statistical image of the probability of each color, and then the distance between the colors by histogram intersection method to measure the similarity between each color histogram, In order to realize the function of image retrieval. Follow-up on the basis of the color histogram derived from the accumulation of color histogram, color moment (generally take the first three moments), color correlation graph and other color features to carry out the image retrieval, here is no longer explained. From the above introduction can be seen in the color characteristics of the main consideration of a variety of color distribution, but these algorithms do not consider the human eye on the distribution in different space, the proportion of the same color difference, for example: a box of the surface of the light surface, when the strong light exposure, Some areas, such as the red marked in the area of the map is difficult to distinguish its true color, and thus lead to the calculation of color histogram and other statistical characteristics of the deviation, resulting in inaccurate search results, and color characteristics of this statistical feature can not exactly give the query image is what the specific object, Only images with the same color distribution can be given, so they are often used as the primary search feature in the retrieval system.

3. Texture features

Texture features mainly describe a certain regularity change of the image pixel grayscale set or color. Texture-based retrieval can usually be divided into two kinds: statistical method and Structure method.

Statistical methods mainly based on the statistical properties of Gray in the image to determine the texture characteristics and the relationship between the characteristics and parameters, at present, Tmaura and other people based on the visual perception of texture and cognition of the texture feature model, more commonly used to the characteristics of roughness, contrast, direction, line image degree, Regularity and coarseness, especially in roughness, contrast and direction of application more widely. Statistical methods are mainly used to analyze fine and irregular objects such as wood grain, sandy land and lawn.

The structure method mainly describes the texture characteristics according to the texture primitives and their arrangement rules. The structural method is to analyze the characteristics of the image with very structural regularity. Suitable for the texture of a class of elements such as cloth or bricks, and the retrieval of objects that arrange the comparison rules.

4. Shape Characteristics

Shape feature is another important feature of image content, which is a basic problem of computer vision and pattern recognition, and there are two main types of shape description methods used in image retrieval: Edge-based and region-based shape method. The former is based on accurate edge detection, using the features of area, perimeter, eccentricity, angle, chain code, interest point, Fourier descriptor and moment descriptor to describe the shape of the object, which is suitable for the image with clear edges and easy access. The main idea of region-based shape feature extraction is to extract the objects of interest in the image by Image segmentation technology, and to extract the image features by the color distribution information of the pixels in the region, which is suitable for the region to be more accurately segmented and the color distribution in the region more evenly. The main difficulty of this method is how to separate the objects of interest of the image efficiently and accurately, which is also one of the hotspots in academia at present.

The shape is usually related to a particular target object in the image, so the shape feature is higher than the color feature and texture characteristics, and the expression of the shape is more complex than the expression of color and texture, in order to get the shape parameters of the target, the first image segmentation/edge extraction, So the shape feature extraction will be affected by image segmentation/edge extraction effect, in the absence of relevant knowledge, the automatic segmentation/edge extraction method is difficult to accurately extract the corresponding target region, and more importantly, the shape of the target is now a very complex problem, People's perception of shape is the result of the combination of retinal sensation and the knowledge of the real world, there is no effective shape model with the subjective sense of the person, besides, the target shape in the image obtained from different angle of view may differ greatly, in order to make the shape match accurately, It is necessary to ensure that the extracted shape features are not affected by the transformation of the image, such as translation, zooming, rotation, etc., which also increases the difficulty of effectively describing the shape, and based on the above reasons, the image retrieval technology based on the shape feature is not particularly mature.

5. Summary

It mainly describes the characteristics of the industry general image retrieval system, which has a common feature is that the extraction speed is fast, the output of image retrieval is a series of TOPN images sorted by similarity from large to small, through the test of some image retrieval engines, Will find that the success rate of the search is not particularly ideal, it is conceivable that if you want to accurately detect the target (TOP1), using the features described above are obviously not feasible, usually consider the use of local characteristics of the better performance of the SIFT algorithm to retrieve, However, the introduction of another problem is that its extraction speed is too slow, it seems to only a few reference images of the image retrieval system, how to ensure the retrieval performance on the basis of improving its feature extraction speed is also a more worthy of further study of the topic. and whether it is color, texture, shape or even sift features do not take into account the image of an important information-spatial distribution, how to integrate this spatial distribution relationship to the existing features is also a worthy of further study.