TWI415032B - Object tracking method - Google Patents

Abstract

The present invention provides an object tracking method, including: firstly, finding an object image area from an initial image defined by a series of continuous images, segmenting the object image area into three object image blocks, and obtaining adaptive histograms thereof; then, according to entropy values of object feature color blocks distributed in the image, dynamically selecting a plurality of image blocks in other continuous images, and also obtaining the adaptive histograms thereof; detecting whether a tracking error occurs, if at least two adjacent image blocks in the continuous images respectively conform to at least two object image blocks, the location of the object in the continuous images can be inferred and tracked. Therefore, the present invention can rapidly obtain image features of a target object, not only the overall computation time can be saved, but also the object moving path can be correctly identified.

Description

Object tracking method

The invention relates to the application of image recognition technology, in particular to an object tracking method.

The technique of object tracking is very broad, and it is most common to track the human body. The conventional techniques of human tracking are mostly directed to the method of figure-ground sementation or temporal correspondence. Most of the contour cut of the object is based on a single image, and attempts to detect the target object in a single image, while the corresponding method of time and space is to try to cut an object into an initial image and try to be in a continuous image. Find out the changes in the same tracker.

The contour cutting method can be further subdivided into five sub-categories: background subtraction, motion-based segmentation, depth-based segmentation, and appearance-based cutting ( Appearance-based segmentation and shape-based segmentation. Taking human tracking technology as an example, most of the methods of human contour cutting are based on some famous detection algorithms or classification algorithms, such as Support vector machine, AdaBoost, etc. method. For example, in 2001, A. Mohan et al. proposed in the IEEE Trans. on Pattern Analysis and Machine Intelligence journals to detect four parts of the human body with a support vector machine (SVM) and then integrate and detect the complete human body with a support vector machine. Algorithm, this algorithm has a good detection effect on the frontal human body, and can reduce the impact of the shadow, but can not track the side of the human body, in addition to using the support vector machine to search the human body part of the image speed is not ideal. In 2004, A. Shanshua et al. proposed a human body detection algorithm based on a single picture in the Intelligent Vehicle Symposium seminar. The algorithm cuts the human body into nine parts and uses the AdaBoost algorithm to detect these parts. In order to detect a person in a single image, but the same as the human body detection algorithm of most single images, it is easy to misjudge the non-human background as a person, and in the case of multiple people, it is impossible to distinguish the different individuals. It is difficult to establish the movement trajectory of each individual.

On the other hand, algorithms based on spatio-temporal correspondence mostly focus on finding the same object in a continuous image. Since different tracking targets in a continuous image have different appearances, and these appearances are often difficult to train in advance, such methods are usually Emphasis is placed on extracting some easy-to-use features immediately, and then detecting this feature in a continuous image to find out the spatial variation of the object in time series. In 2003, K. Nummiaro et al. proposed the most representative tracking algorithm with color histogram as the main feature in Image and Vision Computing. This algorithm uses particle filter to Establishing the spatio-temporal correspondence, this method is widely used in many tracking techniques, but because of the poor ability of color histograms to discriminate objects, other features are often needed.

An example of a color histogram of an image captured by a particle filter using a conventional technique is given below. Please refer to FIG. 1 , which is a gray-scale photographic picture. There are two image blocks in the picture: a face image block 10 and a background image block 20, and the two image blocks are converted into a color histogram by using a particle filter. The color distribution of the two image blocks is counted on the same color histogram by dividing each area into eight color intervals. As shown in FIG. 2, the face image block 10 and the background image block are displayed. The 20 gradation is similar, which causes the pixels in the image to fall into a small number of color intervals. Therefore, the conventional color histogram is difficult to recognize the difference between the two image blocks.

In addition, in 2003, D. Comaniciu et al. proposed a similar practice to K. Nummiaro et al. in IEEE Trans. on Pattern Analysis and Machine Intelligence. This research focuses on the kernel function of color histogram calculation and Calculate the distance function of similarity, and in the part of the establishment of the space-time correspondence, this study is based on mean-shift, because the method used is similar to that proposed by K. Nummiaro et al. The ability to identify objects is poor. In 2004, C. Chang et al. proposed a human cycle motion tracking algorithm based on human body parts in the IEEE Signal Processing Letters journal. This algorithm uses the mean moving integrated particle filter to establish the correlation of human body parts in time series. This algorithm is mainly applied to the tracking of the periodic limb movements on the side, and is not suitable for human tracking in general activities.

In addition, the human body tracking technology also focuses on the human body cutting time and space. In 1997, R. Wren et al. proposed a system called PFinder, which is mainly used to establish a human body tracking interaction with the computer. It is a background model of Gaussian distribution, and the space-time correspondence is to integrate color information and position information into a vector to calculate the similarity of continuous images. The focus of this research is to integrate the structure of human body cutting time and space, as well as the characteristics of use and The way of time and space is so simple that it can only be used in relatively simple scenes. In 2000, I. Haritaoglu et al. proposed an algorithm for tracking human body parts. This algorithm cuts out the human body with a Gaussian distribution and a background model of some rule-building forces, and locates the limbs with the contours of the human body image. Point, each image head positioning point is regarded as a reference point to estimate the position of other parts of the positioning point to establish the temporal and spatial correspondence of the limb, but such an architecture is difficult to cope with more complicated limb changes. In 2007, B. Wu and R. Nevatia published a study in the journal Intell. Jour. of Computer Vision, which proposed a tracking system combining human body part detection and average mobile tracking algorithm. The part of the human body detection is based on AdaBoost. Measuring the human body's head, torso, and foot parts and the intact human body, and then using the spatial relationship of these detection results in the continuous image to establish the change of the displacement in the time series, when the detected human body part is in the continuous image When the spatial relationship deviation is too large, the average moving algorithm is used to adjust the tracking result. Since this algorithm is still mainly based on detection, there is no deep research on the displacement correlation of continuous images, and the mean moving tracking and human body parts The correlation of detection is not obvious, so the resulting tracking trajectory is less able to present the actual human movement pattern.

However, in most prior art, there is less mention of tracking possible errors, how to find the same person after the error occurs. For human body cutting methods, tracking errors may cause fewer problems, but the opposite is The sacrifice of execution speed and the difficulty in ensuring the established tracking trajectory belong to the same person. In 2006, SL Dockstader and NS Imennov proposed to treat the human body as an object connected by 15 endpoints, and use Hidden Markov Model (Hidden Markov Model). To estimate the occurrence of tracking errors, this method introduces the use of time series tracking endpoint changes to detect the occurrence of tracking errors, but unfortunately the human body model of 15 endpoints is difficult to automatically locate from the image, and this The study did not mention how the error should be corrected after the tracking error occurred.

Therefore, in the tracking technology, whether it is the object contour cutting method, the space-time correspondence method, the combination of the two or the tracking error, there is still room for improvement. In view of this, the present invention proposes an object tracking method. An effective improvement method is proposed for the deficiencies of the prior art.

The main object of the present invention is to disclose an object tracking method, which divides an area of a tracking image into a plurality of sub-blocks for respective tracking, and performs an overall error analysis, so that the tracking can still be continued when tracking a small portion of the object. Objects, and can automatically correct errors during the tracking process, improving the accuracy of tracking objects.

Another object of the present invention is to disclose an object tracking method, which converts color features in the captured image block into adaptive histograms, so as to facilitate image color matching between the image blocks, thereby improving the correctness of the tracking object, and Avoid the use of too much detail useless color information and affect the lack of tracking efficiency.

A further object of the present invention is to disclose an object tracking method for adjusting the number of required feature particles according to the disordered distribution of the tracking object in the image when capturing the image data, so as to improve the tracking performance.

In order to achieve the above object, the present invention provides a method for object tracking, which first selects an initial image and a subsequent continuous image of the initial image; then defines an image region of at least one object in the initial image, and cuts the object image region into at least three An object image block, and obtaining a first adaptive histogram representing a distribution of feature patches in each object image block; and subsequently, dynamically selecting a plurality of image blocks in the continuous image according to the disorder of the object image block in the continuous image And taking the second adaptive histogram, wherein the number of image blocks is related to the disorder; comparing the obtained first adaptive histogram with the second adaptive histogram to detect whether there are at least two adjacent ones The feature color block distribution of the image block conforms to the feature color block distribution of the image block of the object; and the displacement path of the object in the continuous image is obtained according to the position of the image block that meets the condition.

The purpose, technical contents, features and effects achieved by the present invention will be more readily understood by the detailed description of the embodiments and the accompanying drawings.

A general tracking algorithm usually consists of two programs: prediction and update. The so-called prediction program mainly predicts the current state based on the behavior of the previous model sample in the system, and the update process uses the current observation data to adjust the detection state of the tracker. The invention provides a method for object tracking, which is constructed under the condition that an object moves gently in a continuous film, and the object to be tracked is positioned by using several image blocks, and then the object is tracked by the prediction and update means. The position in the picture.

In an embodiment of the present invention, human body tracking is taken as an example. Referring to FIG. 1 , first input a plurality of images having a human body image and continuous time, and select one image as an initial image, and the subsequent image of the initial image is a continuous image. , as shown in step S10 in the figure; defining a human body image area in the initial image, and cutting the human body image area into at least three image blocks, hereinafter referred to as human body image blocks, and the present embodiment is distributed in three parts of the human body in three different parts. For example, the human body image block is a head image block, a body image block, and a limb image block, and the first adaptive histogram of the three human body image blocks is taken, as shown in step S12.

The method for obtaining the human body image block is to cut the two cutting lines of the human body by a set of common proportions, and then adjust the cutting line up and down to find the largest difference between the color difference between the head and the upper body or the upper body and the lower body of the human body image, and thus the tangent line In the range of the interval, the image block is obtained according to the position of the center of gravity of the block color, and the image block is subjected to gray scale processing and its adaptive histogram is taken. The adaptive histogram is used to describe the proportion of different colors in the image block, so the adaptive histogram is a representative feature patch distribution in the image block.

In order to find the human body image in the continuous image after the initial image, as shown in step S14, the particle filter is used to dynamically select a plurality of image blocks from the continuous image according to the entropy to slow down the calculation time in time. In particular, the present invention dynamically adjusts the required image blocks according to the disorder, and the so-called chaos is so that the color of each area in the image is very similar to be tracked by the human body, or when the tracked human body is missing, it is difficult to identify When the human body image is taken out, the particle weights obtained in the above two cases are widely distributed. Therefore, the present invention will select a large number of image blocks to enhance the human body image tracking; and the small degree of disorder is only a small part of the image and The human body being tracked is similar, and the particle weight distribution is limited to these few regions, so the chaos entropy value is small. In this case, the invention only needs to select a small number of image blocks to track the human body. In this step, the color block distributions of the selected image blocks are further characterized, and a second adaptive histogram of each image block is taken.

Then, step S16 compares the first adaptive histogram of the three parts of the human body image block of the initial image, which represents the image features of the three parts of the human body, and compares with the second adaptive histogram taken from the continuous image, and Get tracking error detection results. There are three types of tracking error detection results, as described below:

First, the wrong human image block is not detected in the continuous image, that is, the adjacent three image color distributions corresponding to the head image region, the body image region, and the limb image region are found in the continuous image. The two adaptive histograms can determine the image blocks of the three second adaptive histograms, which represent the position of the human body in the continuous image, as shown in step S18.

Second, an erroneous human image block is detected in the continuous image, that is, two adjacent second adaptive histograms are found in the continuous image, which are in accordance with the head image area, the body image area or the limb image area. Any two regions feature color patch distribution, because one of the human body image regions does not appear in the continuous image, the human body may be shielded, and the missing human body image region is estimated based on the detected image position of the human body image. Block position, in other words, if the position of the image area corresponding to the head image area and the body image area is found in the continuous image, the position of the limb image area can be estimated, and the position of the human body image in the continuous image can be known, and the other two cases are the same The position at which the human body moves can be estimated, as shown in step S20.

Third, two or more error image blocks are detected in the continuous image. If the situation occurs, the process of the embodiment returns to step S12 to reselect the current continuous image as the initial image. And find the human body image in the vicinity of the disappearance of the human body image in the previous continuous image, and define the position of the human body to re-track the human body.

Further, in order to increase the recognition ability of distinguishing different objects, the present invention proposes an adaptive histogram to capture the color patch features of the image block. When a particle filter is used, if a large number of particles are extracted, the speed may be lowered. However, if the number of particles is too small, the accuracy may be lowered. Therefore, the present invention dynamically adjusts the number of particles picked up according to the turbulence value of the particle weight. The conventional technique for calculating the color histogram usually divides the color space into a number of small color bins (8×8×8), and when the area of one image block is 32×32, it is divided in each color interval. The expected number of pixels is 2, and the quantized color distribution resolution space needs to be strengthened. The invention increases the color interval classification of the tracking image, selects three independent channels YC _b C _r as the color space, assigns the feature values of each color to the corresponding color interval, and the image pixels in each color interval. The expected number is 128, which more fully expresses the color distribution of the image.

In the color histogram described in the prior art, the face image in FIG. 1 cannot be taken out. In order to solve this problem, the present invention selects a color histogram H to perform histogram equalization to form an adaptive histogram, first in FIG. The face image block 10 is equalized with the color histogram H of the background image block 20. The color histogram after the equalization is z, and the formula is z=M(H), where the M function is the color to be referenced. The histogram H becomes equal to the adaptive histogram z, and the color histogram H' of the other image block is equalized to form another adaptive histogram z'=M'(H). In this way, the pixels of two similar image blocks can be prevented from falling into the same color interval, so that the overall color can be effectively and quickly quantized, as shown in FIG. 4, which is the image block 10 and 20 of FIG. The adaptive histogram is taken out, and compared with the second figure obtained by the prior art, the adaptive histogram of the present invention resolves the finer and more effective color block information of the human face image block 10 and the background image block 20, so that it can be easily distinguished. The difference between the two areas.

Since the tracker continues to move, its appearance may also follow changes. The present invention also takes these external factors into consideration, and the initial adaptive histogram feature value can be updated with each successive image. The function definition is as shown in formula (1):

H _t = H _{t -1} ×0.9+ Q _t ×0.1 (1)

Where H _t and H _{t -1} are the histograms that will occur when time is t and t-1, and Q _t is the color histogram taken directly from time t, where the coefficients 0.9 and 0.1 can be continuous according to the influence The environmental factors of the image change. Therefore, the present invention updates the initial image information as the tracking time changes.

In the particle filter, the particle with local maximum weight is very important for tracking object estimation and particle reset, and a technical feature of the present invention can change the number of particles captured to control the coverage in the state space, so that the maximum The state of the weight (tracking object) is located in the image. When using a particle filter to track an object, if the appearance of many areas is similar to that of the tracking object, the particle weight will approach a uniform distribution, and thus the image has high turbulence. Conditional probability if the object is not tracked It will become lower and its weight distribution will be closer to uniform. Therefore, the object tracking method needs to introduce a larger number of particles. On the other hand, if only one area in the image looks similar to the tracking object, the weight will be concentrated on the particles in this area (the number of particles is small), resulting in low image chaos, in which case only small Object tracking can be performed by image blocks and a small number of particles. The above description shows the relationship between the time t, the disorder and the number of particles in the present invention as shown in the formula (2):

Where ω ^{( n )} is the Nth particle weight, and C is a constant that controls the rate of particle increase. For example, when C is set to 2, the maximum number of particles at t is 2‧N _t-1 . In the present invention, the C value is set to 1.2, and in order to avoid an increase or decrease in the number of particles, the number of N _t is limited to (200, 1000).

Furthermore, in order to reliably track the human body, the present invention divides the human body into a head, a body, and a limb (a part of the two feet) to simultaneously track, and divides the human body in the initial image through an adaptive Gaussian background model, assuming a large The three parts of the part are similar proportions. See Figure 5(a). The present invention sets two horizontal line positions according to a predetermined height ratio to cut out three parts of a human body, and the predetermined ratio is H _{h .} , H _t and H _l . After that, the two horizontal lines are finely adjusted vertically until the horizontal line separates the maximum color difference of the three parts. If the vertical adjustment distance is too large, the original height ratio will be used to track the three parts. The height ratio of the three parts after the fine adjustment is changed to R _h , R _t and R _l . The range of the cut may include the background area and the interference, and the shape of the three parts will change with different tracking characters or different postures. The invention takes a more reduced range of the cut region, which is three human image blocks 30, 32, 34, as shown in Fig. 5(b), the pixel is set to Where (C _x , C _y ) is the center of the rectangle, the center of (C _x , C _y ) is defined as the center of all pixel masses in the R part, and the range covered by the (S _x , S _y ) rectangle is referred to as x And the standard error of the pixels in the y space. Figure 5(b) shows that the three human image blocks 30, 32, and 34 are in the human body position, and the colors inside the three squares are respectively consistent.

Finally, a technical feature proposed by the present invention-tracking error detection is that after dividing the human body into three blocks, the relative positions of the three parts and the moving speed of the parts are limited, but the human body is not steel. The body structure, so the tracked human body image is prone to abnormal deformation, and thus irregular changes occur in the continuous image. Therefore, the present invention uses a support vector machine (SVM) as a classifier to identify whether the human body image block is properly performed. Human body tracking, in which SVM is a common classifier to find the hyperplane in high-dimensional space to distinguish the two species by the largest edge region, and the optimized hyperplane is calculated according to formula (3), formula ( 3) as shown below:

The invention selects the radial basis function as the core function K to draw the feature vector into the high dimensional space, the category label Indicates whether the feature vector is a tracking error. For the subset of training data aggregation, called the support vector, the coefficients α _i and b are determined by solving the quadratic equation problem of large scale.

The present invention designs three SVM error detectors to detect tracking errors of three human image blocks. If the body image blocks of at least two parts are not tracked, the SVM error detector becomes inefficient. Because it is not easy to identify which part is abnormal, the present invention designs a fourth SVM error detector to determine whether the error condition is one or two human image block errors.

When using the SVM error detector, the estimated state of the three parts of the initial image and the feature vector of the state change between the time t ₀ and the time t ₀ -Δt of the previous image are defined as [ RS _H ( t ₀ ) , RS _T ( t ₀ ), RS _L ( t ₀ ), RS _H ( t ₀ )- RS _H ( t ₀ -Δt), RS _T ( t ₀ )- RS _T ( t ₀ -Δt), RS _L ( t ₀ )- RS _L ( t ₀ -Δ t )] ^T , where the vectors RS _H ( t ₀ ), RS _T ( t ₀ ) and RS _L ( t ₀ ) are represented as three human image blocks of time t Relative state vector. The relative state vector is defined as equation (4):

Among them, S _H (t), S _T (t), and S _L (t) are estimated state vectors of the head, the body, and the limbs.

In succession, the particle filter is used to track three body parts in a series of films, and the training samples are manually labeled from the tracking results of each SVM error detector. The present invention uses 150 for each SVM error detector training. The training samples will not cover the image block of the human body, and the state vector will be marked as a negative value. When these image blocks falling inside the human body are marked as positive values, the tracker may be difficult to continue tracking due to tracking errors. The present invention preferentially selects a suitable negative rate, and can adjust the parameters of the SVM error detector to accomplish the target. Therefore, when a tracking error occurs, the SVM error detector must adjust the state of the tracked target. For example, if the body tracking loses two or three parts, the present invention will detect the vicinity of the tracking position of the previous target, and The entire tracking process is initialized; and if only the state of a single part is abnormal, the present invention will utilize the other two parts to adjust the position and size of the abnormal part.

In addition, in the case where the tracking object is obscured in the image, and the appearance of the tracking object may not be correctly taken out from the figure, the present invention proposes to reduce the chance of tracking error caused by image shielding by marking.

There are usually two types of obscuration: two tracking object image overlays, and a background obscuration tracking object. For the former situation, the present invention uses the human body tracking method to predict that two tracking objects may overlap, and when the overlap is about to occur, the technique of tracking the error detection is stopped, and only the original particle filter is used to detect the tracking object. Speed and direction, until the overlap of the two tracking objects disappears, and then use tracking error detection. However, if the background includes a shelter, which may be a pillar or a door, the present invention manually marks the position of the shield in advance, and when the tracer falls within the scope of the shield, only the original particle filter is used to detect the tracer. Speed and direction, stop using the tracking error detection technology in this part to avoid interruption of the wrong body tracking.

In summary, the present invention provides a method for object tracking, which divides object tracking into multiple image blocks, and performs tracking error detection analysis after each tracking, so that the object in the image is partially obscured or distorted. The correct tracking result can still be corrected by tracking error detection. Moreover, in order to efficiently capture the adaptive histogram of the image block, the present invention can highlight the distribution of the feature color patches in the object image block, so that the subsequent comparison process can be performed quickly, and the correctness of the tracking object is improved. In addition, the tracking technology requires a large number of image blocks to be obtained from the continuous image. The present invention uses the dynamic elastic capture quantity to obtain the information required in the continuous image to further improve the performance of the present invention.

The embodiments described above are merely illustrative of the technical spirit and the features of the present invention, and the objects of the present invention can be understood by those skilled in the art, and the scope of the present invention cannot be limited thereto. That is, the equivalent variations or modifications made by the spirit of the present invention should still be included in the scope of the present invention.

10. . . Face image block

20. . . Background image block

30. . . Head image block

32. . . Body image block

34. . . Limb image block

Figure 1 is a schematic diagram of a grayscale photographic image.

Figure 2 is a color histogram taken from two image blocks in the photographic screen.

Figure 3 is a flow chart showing the steps of an embodiment of the present invention.

Figure 4 is an adaptive histogram taken from two image blocks in the photographic picture.

Fig. 5(a) is a schematic view showing the invention in three parts according to a pre-proportional cutting of the human body.

Figure 5(b) is a schematic diagram of three human image blocks taken in the present invention.

Claims (17)

A method for object tracking includes: (a) selecting an initial image and subsequent plurality of consecutive images; (b) defining an object image region of at least one object in the initial image, and cutting the object image region into at least three Object image blocks, and respectively obtaining first adaptive histograms of the image blocks of the objects, wherein each of the first adaptive histograms represents a characteristic patch distribution of the corresponding image block of the object; (c) from the continuous A plurality of image blocks are dynamically selected by a particle filter in the image, and a second adaptive histogram of the image blocks is obtained, wherein the number of the image blocks is determined by the disorder of each successive image, and each The second adaptive histogram respectively represents the characteristic patch distribution of the corresponding image block; (d) the first adaptive histogram and the second adaptive histogram are compared, and the continuous image is known Whether the feature patch distribution of the at least two adjacent image blocks respectively conforms to the characteristic patch distribution of the image blocks of the objects; and (e) according to the image blocks conforming to the objects The positions of the image block of the intrinsic color distribution of the object to obtain a continuous image at that location, the movement of the object to track. The method of object tracking according to claim 1, wherein the continuous image is a series of time-continuous images. The method of object tracking according to claim 1, wherein the object system is a human body. The method of object tracking according to claim 3, wherein the object image block comprises a head image block, a body image block or a limb image block to track the head, body or limb of the human body. The method of object tracking according to claim 1, wherein the first adaptive histogram or the second adaptive histogram is based on a color distribution of the object image block or the image block corresponding thereto. A representative feature patch map is taken out. The method of object tracking according to claim 1, wherein in the step (d), the continuous image has two adjacent adaptive histograms of the image blocks, which respectively conform to the image of the objects Two of the first adaptive histograms of the square. The method of object tracking according to claim 6, wherein the step (e) is to obtain the position of the object in the continuous image according to the two image blocks that match the image block of the object. The method of object tracking according to claim 1, wherein in the step (d), the continuous image has three adjacent adaptive histograms of the image block, which respectively conform to the image of the object The first adaptive histogram of the square. The method of object tracking according to claim 8, wherein in the step (e), the object is located at the position of the continuous image according to the three image blocks that match the image block of the object. The method for tracking an object according to claim 1 further includes a determining step of determining, by the training result of the support vector machine algorithm, the correctness of the object at the position of the continuous image. The method of object tracking according to claim 1, wherein the continuous image has at least one mask, and the position of the mask is marked as an error detection range, and the object is about to fall into the error detection range. When you stop tracking. The method of object tracking as described in claim 1, wherein the initial image is determined The object image area of the two objects. The method of object tracking according to claim 12, wherein the tracking of the object image areas is about to overlap. The method of object tracking according to claim 1, wherein the first adaptive histograms are weighted to add the second adaptive histogram having an object feature patch distribution, thereby generating a plurality of new histograms. The first adaptive histogram causes the first adaptive histograms to be updated over time. The method of object tracking according to claim 1, wherein the disorder is related to the appearance of the object and the appearance of each part of the continuous image, the large degree of disturbance representing the appearance of the object and the continuous image. Many of the areas have similar appearances, and the small degree of spoilage means that the appearance of the object is similar to the appearance of a small portion of the continuous image. The method for object tracking according to claim 15, wherein in the step (c), when the disorder is large, the number of the image blocks is extracted. The method of object tracking according to claim 15, wherein in the step (c), when the disorder is small, the number of the image blocks is reduced.