Wear dyed beginning of the heart - computer vision status (3.1): image segmentation
In order to pass the material, shape, structure recognition target, the image segmentation should be an inevitable choice . At this point, computer vision researchers be barely reached a consensus, image segmentation even been listed as one of the basic tasks of computer vision many researchers. Many image segmentation algorithm was put out, it can be grouped into a method based on threshold, clustering based, region-based method, a method based on graph cuts, and activity level set boundary model, based on convolutional neural network method .
The method based on the threshold value, there is a global threshold value, but also to use different thresholds for different regions of the image. The key method is to determine such a threshold value, but for most images even through all values achieved can not find a reasonable segmentation threshold , so the algorithm is commonly used in character recognition, fingerprint recognition, and remote sensing various split index images ( water index, vegetation index, drought index, thermal infrared images).
Those who seriously think about or try it for yourself, will admit unacceptable result clustering based segmentation method. Clustering algorithms are all directly or indirectly assumed sample density sample density than the interior of the category boundary of the category . After some of the histogram of an image observed will find the pixel value of the image element shape of the distribution space are usually similar to a bell curve (only one projection), that is to say in accordance with the clustering algorithm should assume all the pixels of a poly class . Even simultaneous use of multiple image channels, there is no change this situation. No or difficult to use space-based clustering information directly to defects, resulting in the segmentation result of complex shape, apparent boundary crossing . Based on Clustering generally are other algorithms as a preprocessing step, in the field of image processing for compressing the number of colors.
Based clustering method has the basic superpixel segmentation algorithm clustering algorithm called SLIC is quite different, in the first image uniformly sprinkle seed point, and then performs clustering KMeans at a constant radius, as used the distance measurement method to consider spatial information. The algorithm has great advantages over square grid-shaped divided super pixels, but due to the natural defects of clustering algorithms for segmentation will lead to super-pixel still evident across the border. If the goal is to maintain the accuracy of shape, SLIC is not a good pre-selection .
Region-based method currently contains seeded region growing, region splitting and merging, watershed. The key is that the seeded region growing seed point selection, the measure of similarity, growth regulation, wherein the similarity measure is the key. Commonly used methods similar measure pixel value of the color, texture, similarity measure is the most crucial, regional division and merger law. Regional split and merge Another key is split and merge mechanisms. Watershed algorithm is typically used in the gradient map, the gradient value as an elevation value, contemplated gushing out from the local minimum value of the height, the water levels of the whole image at the same rate, two meet at the gushing the construction of the dam. When the water level rises to the highest point, all the pixels have been submerged, all the dams forming the image segmentation. Watershed algorithm faint edge also has a very good response , and consequently lead to over-segmentation. To maintain the shape, the watershed region split and merge should add there is a large room for improvement .
Dividing an image conversion method is based on graph cuts FIG minimum cut problem, the key is the right side of the weight set, which is just another name similarity measure . And therefore did not grow full image segmentation method based on graph cuts than seed, regional split and merge better; furthermore minimum cut problem is NP-hard optimal solution. FIG method worth mentioning is cut based GrabCut Graphcut and, prior to the segmentation algorithm for the background, the need for manual tagging rough background and foreground portions. Direct significance of these two algorithms is that the prospects for interactive image processing to extract.
Useful model and moving boundary lies in the same horizontal set of interactive foreground extraction , it is necessary to provide artificial initial foreground range of the curve, and range of the curve so that under the effect of the evolution of the energy function obtained from the image data, so as to gradually approach the edge of the foreground, finally found outlook edge when it reaches the edge of the prospect minimum value of the energy function. Moving boundary model directly on the curve range evolution, and level set according to the constructed three-dimensional surface image data of the initial range of the curve and, indirectly through evolution Evolution range of the curve and how to take three-dimensional surface contours. At the front edge of a certain background has a larger gradient, so the most important structural basis is the image gradient energy function, the gradient will only rely on local optimization, will join the general curvature of the curve and the other a priori knowledge. Model and moving boundary level set method is desirable, but the role of passing through curves and surfaces as well as other local maximum gradient prior to break , and in the foreground segmentation is not better than the methods based on graph cuts, is true.
Convolutional neural network-based methods include semantic segmentation, segmentation example, three kinds of panoramic division, in the final analysis and a range of pixels within the pixel areas of the establishment or class instance mapped to a known sample data. Thanks to the advantages of neural network can simulate any function under conditions of a large number of sample data involved in the training of the established out of the mapping model with high accuracy. However, defect mapping model of the machine is obvious, migration (can split the network can not be separated cat dog) and generalization (not only split white network segmentation cat) almost none. In addition, as overly complex mapping model, a slight change in the input data can lead to unpredictable errors (pixels attacks) .
These are all developments of image segmentation, who had to try people will admit that even human beings are difficult to extract the shape and structure from the results of the current image segmentation feature. Ask, in the absence of identified targets, can be divided in the end to what extent? Currently image segmentation algorithm still be able to continue to improve ? I personally feel that the current progress has not yet reached the limits of image segmentation, although impossible to achieve human level, but there is still much room for improvement. The key question is what direction to ascend, that is what should be used to enhance the image segmentation indicators to assess the ability of ?
Currently evaluation image segmentation algorithm is based on the IOU, IOU is emphasized that the direct area, while the area is very poor discrimination features of the polygon is . For many natural target for complex shapes, but the shape and structure of large IOU is a far cry, so based on the evaluation index IOU's of no value to emphasize the shape and structure. But there is no description of the shape and structure feature, as if caught in a dilemma. In fact, not difficult to see, if the border is accurate, then connect the exact boundary can form an accurate segmentation. A small amount of dislocation boundaries is inevitable, so when evaluating the accuracy of the border need to allow a certain degree of misalignment. But Ruoguo border are missing, the case can not recover the shape and structure. So far, an evaluation of segmentation algorithms ready to come - the boundary miss rate. Another evaluation is the corresponding - border redundancy rate .