Visual error correction of continuous aerobics action images based on graph difference function

In order to realise the correction of visual errors of continuous aerobics action images, the acquired images were first processed by holographic infrared vision scanning technology [5] and the imaging projection of continuous aerobics action images was realised by image space scanning method, and then the template feature combination equation of three-dimensional imaging of aerobics action images was calculated: (1) I(x,y) = h(x,y)*f(x,y) + δ(x,y) I(x,y) = h(x,y)*f(x,y) + \delta (x,y) In Equation (1), H (x,y) is used to describe the parallax function; * is used to describe convolution; F (x,y) is used to describe the edge contour pixels of continuous aerobics action images; δ(x,y) is used to describe the sub-pixel features of aerobics action images.

The parallax function of the image pixel set was registered according to the visual error, and then the sub-pixel template matching of the image was achieved through vector quantisation technology [6], and the pixel output after template matching was obtained: (2) I(x,y)=f(x,y)+δ(x,y) I(x,y) = f(x,y) + \delta (x,y) The estimated pixel value of the edge contour of the continuous aerobics action image is obtained: (3) f∧(x,y)=αF(x,y)+(1−α)A+σ2 f \wedge (x,y) = \alpha F(x,y) + (1 - \alpha )A + {\sigma ^2} In Equation (3), F(x,y) is used to describe the statistical features of strong texture set parameters obtained after the holographic infrared vision scanning projection of continuous aerobics action images. α is used to describe the background differential pixel characteristic factor; A describes the background differential pixel distribution texture set; σ² is used to describe local variances.

Holographic projection technology is used to obtain continuous calisthenics motion images, which can provide reliable original images for visual error analysis and correction.

Fractal coding processing was carried out on the processed continuous aerobics action images to provide a basis for visual error correction [7]. For the aerobics action image I(x, y) with the size of M × M, the coded R block size is m × m, the number of R blocks required to encode is MR = (M/m)², the size of D block in the code book Φ is 2 m × 2 m, Φ represents code book capacity [8], Φ = 8[(M − 2 m)/λ + 1]², where λ is used to describe the code book step size. If the best matching block of each R block is obtained according to the global search method, the whole codebook needs to be traversed, MRD = MRΦ matches are required in the coding process; due to the large size of code book capacity, the computational complexity will be increased and the efficiency will be affected, so the search space needs to be reduced [9]. The image block features are defined next and the full search at matching time is transformed into neighbourhood search by the features [10]. First, the image block U = [UI,j] ∈ Rm × m is reduced to an image block U = [UI,j] ∈ Rm/2 × m/2 by means of the mean value of four neighbourhood pixels, the calculation formula for self-similar features of image blocks is as follows: (4) W(U)=‖V¯−v¯•I‖/‖V−v¯•I‖ {\rm{W(U)}} = \left\| {\overline V - \overline v \bullet I} \right\|/\left\| {V - \overline v \bullet I} \right\| In Equation (4), V is used to describe image blocks, V̅ is used to describe a shrinking image block, is used to describe the 2-norm of V̅ without the mean, is used to describe the 2-norm of V without the mean. Under the condition of high similarity between R block and D block, they can form a matching pair.

The self-similar eigenvalues are arranged in order from small to large to complete the sorting of block D in the code book, then, in the ascending code book Φ, the binary method is used to search the initial matching block D with the smallest difference from the self-similar eigenvalue of W (R) the coding block R: (5) Dinit = {D∈ΦminW(R)−W(D)} Dinit = \{ D \in \Phi \min W(R) - W(D)\} Coding search is completed in the k neighbourhood of Dinit. For each coded R block, search codebook ΦW can be obtained through Equation (6): (6) ΦW = {D∈Φ,D∈M(Dinit,k)} \Phi W = \{ D \in \Phi ,D \in M(Dinit,k)\} The D block corresponding to the minimum matching error of ΦW and R block is obtained, and the fractal code of R block is obtained, as ΦW capacity is lower than that of bit capacity, it can effectively improve the encoding speed [11].

There is inevitably a certain error between the original continuous calisthenics action image I and visual image I∧, that is (7) I=I^+E I = \hat I + E In Equation (7), E is used to describe the error image. The interpolation operator g is applied to both sides of Equation (7), so (8) IK=g(I^+E)=g(I^)+g(e)=I^K+ek {I_K} = g(\hat I + E) = g(\hat I) + g(e) = {\hat I_K} + {e_k} In Equation (8), is used to describe the fractal partial error of local self-similarity, is used to describe the error between and its estimate, at the same time, it can be estimated by Equation (9) (9) e^k=gk(e)≈ek {\hat e_k} = {g_k}(e) \approx {e_k} In Equation (9), g_b represents the two-cube interpolation operator. In order to reduce the interpolation error, is added to as compensation, and it is taken as the optimal estimate of.

On the basis of the above analysis, the visual error correction process of continuous aerobics action images is given: (1)

Establish the covering set {Si} of continuous aerobics action images, each subblock Si has size m₀ × m₀, S_j is used to describe similar aerobics blocks, S_i ∩ S_j = Φ at the same time.

Stereoscopic matching error detection and correction [12] can be proved by a three-dimensional image of a pentagon object. The boundary can be obtained by the linear feature extraction system, and the parallax at each boundary point can be obtained by the Intra-Scanline Matching Algorithm. For display, the obtained parallax is interpolated along the polarisation line to generate a dense parallax graph.

Since the parallax of the line segments along the line varies linearly and if all the boundaries match correctly, almost all the marked visual impairments will fall on a straight line, as shown in Figure 1. This is not the case, as shown in Figure 2.

Fig. 1

Fig. 2

If the line segment can be interpolated with the data corresponding to the correct parallax, we can not only detect and correct the false match, but also fill the parallax information in the unmatched part and the occluded part. Before interpolating the parallax value, the marked point has to be discarded due to an incorrect match. If more lines match the correct line segment than a single wrong line segment, that is, the local error, then thin strips in all directions and positions can be fitted to the marked parallax length data to detect the error. According to the definition of local error, it can be concluded that the thin band with the largest number of points has the parallax value corresponding to the correct match, and the rest are wrong. The use of thin strips instead of straight lines is due to the discrete location of the boundary and the error at the boundary position so that all correctly matched points do not completely fall on a line.

By using the improved Hough transform technique, we can not only attach the thin tape to the parallax data, but also correct the parallax of the horizontal line segment which is parallel to the image plane and the non-horizontal line segment.

Based on the stereo algorithm of typical boundary points, we use dynamic programming to calculate the best matched boundary points along the polarisation line, that is, the minimum total value with the matched boundary pair. Thus, it is possible to compare the performance of four different matching valuation functions of Cost Functions (C.F.). These functions (except the first one) do not correspond to any standard stereo matching algorithm and have been chosen only to prove such a valuation algorithm.

Tables 1 and 2 quantitatively compare the performance of the dynamic programme-based matching algorithms when testing two stereo image pairs using four different matching graph phase difference functions. The calculation methods and steps used in the measurement process are as follows: (1)

Total number of line segments selected and use line segments greater than the fixed threshold to complete error detection;

The total number of line segments processed: according to the selected line segments, the line segments with more than a fixed number of matching boundary points and the correct matching are used for error statistics and correction to ensure their reliability;

Percentage of corrected boundary points: it is either the number of corrected boundary points or the number of filled boundary points according to processing all boundary points on the line segment to make their parallax. We analyse and compare the errors generated by the following four matching valuation algorithms:

In Tables 1 and 2, the performance of four algorithms for image processing of two complicated three-dimensional objects is listed. From the use of these four matching algorithms on the two landscape quantitative calculations, it can be seen that the performance of the graph difference function I and the graph difference function is very similar, and the graph difference function of the sum and the graph difference function of the performance is very different. In order to obtain the stability of the performance test of this evaluation algorithm, various stereo image sum algorithms must be considered. Comparable values of the two different scenes in this paper are listed in the table.