JPH05314253A - Image converter - Google Patents

Abstract

PURPOSE:To provide a luminance converter for improving memory efficiency and edge detecting accuracy. CONSTITUTION:At the image converter to convert the luminance of an image to a density value based on a prescribed characteristic and to prepare the image expressed with the density value in an image memory, a converting means is provided to convert the luminance to a density value L corresponding to a characteristic [log (s) + B] almost proportional to the logarithm of the luminance in an area where luminance (s) of the image is comparatively large (s >= a) and to convert the luminance to the density value corresponding to a characteristic (As/a) almost proportional to the luminance in an area where the luminance (s) of the image is comparatively small (S<a.) Since the quantizing error of the luminance is fixed at a/2A in the low luminance area and can be enlarged rather than the accuracy of the own luminance, the bits of the memory are effectively used. Further, since the weight of differential arithmetic is reduced in the low luminance area concerning the density converted by the characteristic, the edge detecting accuracy is improved.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image conversion apparatus for converting the brightness of an image according to a certain characteristic, and more particularly, to improving the memory efficiency when storing an image having a wide dynamic range in an image memory and improving the edge detection accuracy of the image. The present invention relates to an improved device.

[0002]

2. Description of the Related Art In the case of detecting the boundary lines (edges) of the areas where the reflectances of the object to be imaged are r1 and r2 on an image plane obtained by imaging the object, by image processing,
The following method is adopted as an example. It is assumed that the density value in a grayscale image obtained by picking up an image of an object is almost proportional to the brightness of the object. A grayscale edge image is obtained by applying a differential operator (Sobel operator, etc.) based on the spatial differential idea to the grayscale image. Various difference type operators are defined by various difference arithmetic expressions regarding the density values of adjacent pixels.

Fundamentally, the above-mentioned edge detection is I (r
2-r1) is calculated. However, I represents the intensity of the illumination light. In this method, the edge strength I (r2-r1)
Is proportional to the illumination intensity I. Therefore, when the illumination intensity I is non-uniform, the edge strength is small in a place where the illumination intensity is small and is large in a place where the illumination intensity is large. In general,
After the grayscale edge image is obtained, the process of obtaining a binarized edge image by binarization is executed. However, when the illumination intensity I is not uniform, the edge intensity depends on the illumination intensity, and it is difficult to obtain a highly accurate binary edge image.

Regions r1 and r2 are independent of the illumination intensity.
As a method of obtaining the edge between
There is a method of detecting an edge by paying attention to the ratio of the ratio and r2 / r1, for example. This method has a problem that when an image whose density value is proportional to the brightness of the object to be imaged is used, it is necessary to apply an operator using division, resulting in a large amount of calculation.

On the other hand,

Since log (Ir2) -log (Ir1) = log (r2 / r1), the density values Ir1 and Ir2 are logarithmically converted to obtain a logarithmic image, and a differential operator is applied to this logarithmic image. By doing so, the edge image corresponding to the ratio r2 / r1 can be obtained without using division. That is, an image that has been logarithmically converted so that the density value is proportional to the logarithm of the brightness of the object is stored in an image memory, and a differential operator such as a Sobel operator is applied to this image memory to obtain the illumination intensity Even if it is non-uniform, it is possible to obtain an edge image with the same amount of calculation as the conventional method.

By the way, generally, when the intensity of the illumination light on the object to be imaged is extremely uneven, the dynamic range of the brightness of the object to be imaged becomes wide. When the dynamic range of luminance is wide, converting the image of the imaged object so that the density value is proportional to the logarithm of the imaged object's brightness and storing it in the image memory is as follows. Image signal having a wide dynamic range from high luminance to low luminance can be stored in the image memory.

[0007]

However, as a result of consideration, it was found that the conventional method has the following problems. When the image of the object is digitally stored in the image memory by logarithmic conversion so that the density value is proportional to the logarithm of the brightness of the object, the density value stored in the image memory is quantized. The brightness of the object to be imaged corresponding to this is also quantized. When a is a positive real number, the logarithmic function log a
In the graph, since the gradient becomes larger as a approaches 0, the luminance quantization error decreases as the luminance value decreases when the luminance is logarithmically converted and stored in the image memory.

On the other hand, the image stored in the image memory is obtained by picking up an image with a television camera, for example. An imaging device such as a television camera is a device that converts the brightness of an object to be captured into an electric signal having an intensity proportional thereto by using the principle of photoelectric conversion. Therefore, it is considered that the signal corresponding to the brightness output from the image pickup device includes an error due to a cause such as a noise component of the image pickup device. This error will be called the error of the brightness itself or the accuracy of the brightness itself. In a logarithmically transformed image, there is no problem because the quantization error of the brightness in the image memory is larger than the accuracy of the brightness itself in the high brightness area, but in the low brightness area, the brightness in the image memory There is a problem that the capacity of the image memory is not used effectively because the quantization error of is significantly smaller than the accuracy of the brightness value itself.

Further, when the image is stored in the image memory and the edge detection is performed by the differential operator as described above, the difference between the density values in the part having a large density value (the part having a high brightness) and the density value The difference between the density values in the small portion (the portion with low brightness) has the same weight. However, in general, since the S / N ratio of the density value is lower in the part having the smaller density value, the accuracy of the edge detection result of the part having the smaller density value is relatively deteriorated, so that the edge may not be properly detected. .. In order to avoid this problem, it is possible to cut the part with low density value, but there is a case that the edge cannot be detected well due to the fact that the information of the part with low density value is dropped. There's a problem.

[0010]

The structure of the invention for solving the above-mentioned problems is to convert the brightness of an image into a density value based on a predetermined characteristic, and create an image represented by the density value in an image memory. In the image conversion device that performs the conversion, the luminance is converted into the density value by the characteristic almost proportional to the logarithm of the luminance in the area where the luminance of the image is relatively large, and the characteristic is almost proportional to the luminance in the area where the luminance of the image is relatively small. It is characterized in that it has a conversion means for converting the brightness into a density value, and stores the density value output from the conversion means in an image memory to create an image.

[0011]

In the region where the brightness of the object to be imaged is relatively high, the conversion means converts the brightness into a density value by a characteristic that is substantially proportional to the logarithm of the brightness, and in the region where the brightness is relatively low, the brightness becomes the brightness. It is converted into a density value by the characteristic that is almost proportional.

[0012]

Since the present invention has the above-mentioned converting means, the quantization error of luminance is not so much reduced in the region of relatively small luminance as compared with the region of relatively large luminance. When storing an image with a wide dynamic range, the capacity of the image memory can be saved, that is, the memory efficiency can be improved. Further, since the quantization error is substantially uniformed over the entire range, the accuracy of edge detection is improved when the object to be imaged is unevenly illuminated.

[0013]

EXAMPLES The present invention will be described below based on a specific example. The brightness s of the first embodiment is calculated by the following function,

[0014]

L (s) = A · log (s) + B (s ≧ a) L (s) = A / a · s (s <a) is used to convert to the density value L (s). , Concentration value L (s)
The case of quantizing and storing in the image memory will be described. However, log is a natural logarithmic function, and a and A are constants larger than 0. In addition, the brightness s has a positive value. B is an offset value, and B can be determined so that L (s) is continuous at s = a. That is,

[0015]

## EQU3 ## Using the condition of A.log (a) + B = A / a.a, B = A (1-log (a)). Therefore,

[0016]

## EQU00004 ## L (s) = A.log (s) + A (1-log (a)) (s ≧ a) L (s) = A / a.s (s <a). Also, the first derivative of L (s) is

[0017]

DL (s) / ds = A / s (s ≧ a) Since dL (s) / ds = A / a (s <a), L (s) is up to the first derivative at s = a. It is continuous. The constants a and A are, as described later,
It is a constant that can be determined mainly according to the accuracy of the brightness itself.

First, an image with a wide dynamic range is
Explain that memory can be efficiently used and stored. [When logarithmic conversion is performed on the entire area] The brightness s
Will be logarithmically converted by the function A · log (s) + B and the density value will be stored as a digital value in the image memory. The density value of the image memory is quantized, and its quantization unit is generally 1. Therefore, since the luminance quantization unit Δs is the amount of change in luminance corresponding to a change in the density value of 1,

[0019]

## EQU00006 ## {A.log (s + .DELTA.s) + B}-{A.log (s) + B} = 1 Therefore,

## EQU00007 ## .DELTA.s = s (exp (1 / A) -1).

As can be seen from the above equation, since Δs is proportional to the brightness s, Δs becomes smaller as the brightness s becomes smaller, and as shown in FIG. Less than the precision of. If Δs is smaller than the accuracy of the brightness s itself, the density value of the memory is redundantly used. On the other hand, if it is attempted to store Δs in a range that is equal to or higher than the accuracy of the brightness s itself, it becomes impossible to store a portion having a small s in the memory, and it becomes impossible to store an image having a wide dynamic range.

On the other hand, in the present invention, the density value is set to be substantially proportional to the luminance in the portion where the luminance s is smaller than a certain level, so that Δs becomes a substantially constant value in the portion where s is small and the portion where s is small. But you don't have to use extra density values. Therefore, it becomes possible to efficiently use the memory to store an image over a wide range of the value of s.

[Conversion Characteristics According to Embodiment] In this embodiment,
By converting the luminance s using the above-mentioned function L (s), when s is larger than a, that is, s ≧ a,

L (s) = Alog (s) + B (s ≧ a), and the density value becomes proportional to the logarithm of the brightness s, but the part where s is smaller than a, that is, s <a in case of,

L (s) = A / a · s (s <a), and the density value becomes proportional to the luminance s.

The luminance quantization unit Δs is as described above.
Since it is the change amount of the brightness corresponding to the change of the density value of 1,

[Equation 10] L (s + Δs) −L (s) = 1 Therefore,

Δs = s (exp (1 / A) −1) (s ≧ a) Δs = a / A (s <a).

When s ≧ a, s (exp (1 / A) −
1)> a / A is mathematically established. Therefore, FIG.
As shown in, in the part where s is smaller than a, Δs is a constant a.
/ A, and Δs becomes larger than the constant a / A in the portion where s is a or more. Since the luminance quantization unit is Δs, the luminance quantization error is Δs / 2. Therefore, if the constants a and A are set so that a / 2A is about the accuracy of the brightness s itself, the quantization error of the brightness in the image memory will be significantly smaller than the accuracy of the brightness itself. Since it disappears, it becomes possible to use the memory efficiently.

[Case of Detecting Edge from Converted Density Value] Next, a case of detecting an edge by a differential operator will be described. [When logarithmic conversion is performed over the entire area]
Consider a case in which is logarithmically converted by the function A · log (s) + B and digitally stored in the image memory. In this case, the difference between the density values of the pixel corresponding to the brightness s ₂ and the pixel corresponding to the brightness s ₁ is

[Formula 12] {A · log (s ₂ ) + B} − {A · log (s ₁ ) + B} = A · log (s ₂ / s ₁ ).

If the accuracy of the brightness itself is δ, then s ₁ >> δ,
In the case of s ₂ >> δ, the difference value is considered to have sufficient accuracy, but the smaller the value of s ₁ or s ₂ , the more
That is, the smaller the density value, the worse the accuracy of the difference value. As an example, s ₁ = 1000, s ₂ = 1500, s
₃ = 100, s ₄ = 150, s ₅ = 10, s ₆ = 15
If the accuracy of the brightness itself is ± 1, then A = 25
Then, as shown in FIG.

[0027]

[Equation 13] A · log (s ₂ / s ₁ ) ≈10 ± 0.04 A · log (s ₄ / s ₃ ) ≈10 ± 0.4 A · log (s ₆ / s ₅ ) ≈10 ± 4 Becomes

However, here, the quantization error when storing in the image memory is ignored. As described above, there is a problem in that the values obtained as a result of the difference have the same value regardless of the difference in accuracy, and the subsequent processing is treated in the same way. For example, when binarization is performed in the next process, there is a problem that the edge binarization result becomes noisy due to the influence of the value of the inaccurate portion. The luminance quantization unit Δs when stored in the image memory is Δs = s (exp (1 / A) −1) as described above, and thus the quantization error Δs / 2 of each luminance is calculated. Then, s ₁ = 1000 ±
20, s ₂ = 1500 ± 31, s ₃ = 100 ± 2, s ₄
= 150 ± 3, s ₅ = 10 ± 0.2, s ₆ = 15 ± 0.
It becomes 3. Therefore, when the effect of this quantization error is calculated and added to the accuracy of the difference,

[0029]

[Equation 14] A · log (s ₂ / s ₁ ) ≈10 ± 0.04 ± 1 A · log (s ₄ / s ₃ ) ≈10 ± 0.4 ± 1 A · log (s ₆ / s ₅ ) ≈10 ± 4 ± 1.

[Case of conversion characteristics according to the embodiment] On the other hand, in the case of the present embodiment, the difference in density value between the pixel corresponding to the brightness s ₂ and the pixel corresponding to the brightness s ₁ is s ₁ ≧ a , S
_{If 2} ≥ a, then

L (s ₂ ) −L (s ₁ ) = A · log (s ₂ / s ₁ ). Further, when s ₁ <a and s ₂ <a,

L (s ₂ ) −L (s ₁ ) = A / a · (s ₂ −s ₁ ) Further, in the case of s ₁ <a, s ₂ ≧ a,

L (s ₂ ) −L (s ₁ ) = A · log (s ₂ ) + A (1−log (a)) −A / a · s ₁ Also, s ₁ ≧ a, s ₂ <a In Case of,

L (s ₂ ) −L (s ₁ ) = A / a · s ₂ −A · log (s ₁ ) −A (1-log (a)).

Now, s ₁ ≧ a, s ₂ = λs ₁ ≧ a, s ₃
≦ a, s ₄ = λs ₃ ≧ a, s ₅ <a, s ₆ = λs ₅ <
When a, λ ≧ 1, A> 0, a> 0, and s _{1 to} s ₆ ≧ 0.

(19) 0 ≦ A / a · (s ₆ −s ₅ ) ≦ A · log (s ₄ ) + A (1-log (a)) − A / a · s ₃ ≦ A · log (s ₂ / s The relationship ₁ ) is mathematically established. Therefore,

The relationship of 0 ≦ L (s ₆ ) −L (s ₅ ) ≦ L (s ₄ ) −L (s ₃ ) ≦ L (s ₂ ) −L (s ₁ ) holds. Therefore, the value of the difference in the low luminance part is less weighted than the difference in the high luminance part.

Further, s ₁ ≧ a, s ₂ = λs ₁ ≧ a, s ₃
≧ a, s ₄ = λs ₃ ≦ a, s ₅ <a, s ₆ = λs ₅ <
In the case of a, λ ≦ 1, A> 0, a> 0, and s _{1 to} s ₆ ≧ 0.

[Expression 21] 0 ≧ A / a · (s ₆ −s ₅ ) ≧ A · log (s ₄ ) + A (1-log (a)) − A / a · s ₃ ≧ A · log (s ₂ / s _Since the relationship of ₁ ) is mathematically established,

The relationship of 0 ≧ L (s ₆ ) −L (s ₅ ) ≧ L (s ₄ ) −L (s ₃ ) ≧ L (s ₂ ) −L (s ₁ ) is established. Therefore, the value of the difference in the low luminance part is less weighted than the difference in the high luminance part.

For the above reason, there is an effect of reducing the weighting of the difference value in the portion where the density value is relatively low and the density value is small, so that the above-mentioned problem is solved.

As an example, s ₁ = 1000, s ₂ = 15
00, s ₃ = 18, s ₄ = 27, s ₅ = 10, s ₆ = 1
In FIG. 5, the difference value and its accuracy are shown in the case where the accuracy of the luminance itself is ± 1. Now decide A = 25.
Referring to the above-described method of defining a, the luminance quantization error a / 2A in the low luminance part is 1/2 of the precision of the luminance itself.
Then, a is determined so that, and a = 25. Then, Fig. 4
As shown in

[0035]

L (s ₂ ) −L (s ₁ ) ≈10 ± 0.04 L (s ₄ ) −L (s ₃ ) ≈9 ± 2 L (s ₆ ) −L (s ₅ ) ≈5 ± It can be seen that the value becomes 2, and the less accurate the part, the less the weighting.

However, the quantization error when storing in the image memory was ignored. The luminance quantization unit Δs when stored in the image memory is, as described above,

Δs = s (exp (1 / A) −1) (s ≧ a) Δs = a / A (s <a) Therefore, when the quantization error Δs / 2 of each luminance is calculated, s ₁ = 1000 ± 20, s ₂ = 1500 ± 31, s
₃ = 18 ± 0.5, s ₄ = 27 ± 0.6, s ₅ = 10 ±
0.5, s ₆ = 15 ± 0.5. Therefore, when the effect of this quantization error is calculated and added to the accuracy of the difference,

[0037]

L (s ₂ ) −L (s ₁ ) ≈10 ± 0.04 ± 1 L (s ₄ ) −L (s ₃ ) ≈9 ± 2 ± 1 L (s ₆ ) −L (s ₅ ) ≈5 ± 2 ± 1.

Second Embodiment As a second embodiment of the present invention, the brightness s is expressed by the function L (s) =
A case will be described in which conversion is performed using A · log (s + a) + B, and the value of L (s) is quantized and stored in the image memory. However, log is a natural logarithmic function, and a and A are constants larger than 0. Also, s is assumed to take a positive value. B
Is an offset value, and since a memory generally stores a value of 0 or more, B can be determined so that the minimum value of L (s) becomes 0. That is, when the condition of L (0) = 0 is used, B = −A · log (a). Therefore,

[0039]

L (s) = A · log (s + a) + B = A · log (s + a) −A · log (a). Note that the constants a and A are constants that can be determined according to the accuracy of the brightness itself and the accuracy of the required brightness ratio, as described later.

First, an image with a wide dynamic range is
Explain that memory can be efficiently used and stored. In the present embodiment, the brightness s is expressed by the function L (s) = A
By converting using log (s + a) + B, when s is sufficiently larger than a, that is, s >> a,

[Equation 27] L (s) ≈A · log (s) + B = A · log (s) −A · log (a), and as shown in FIG. 5, the density value is almost proportional to the logarithm of the luminance s. Like On the other hand, when s is sufficiently smaller than a, that is, when s << a,

[0041]

L (s) ≅Alog (a) + A / as + B = A / as, and the density value becomes almost proportional to the luminance s, as shown in FIG. Also, from a> 0 and A> 0,

## EQU29 ## Since A.log (s) -A.log (a) <A.log (s + a) -A.log (a) <A / a.s (where s> 0), the above 2 The middle part of the two parts has these middle properties.

The density value of the digital image memory is quantized, and its quantization unit is generally 1. Therefore, since the luminance quantization unit Δs is the amount of change in luminance corresponding to a change in the density value of 1,

(30) L (s + Δs) −L (s) = 1 Therefore,

Δs = (s + a) (exp (1 / A) −1). Therefore, as shown in FIG. 6, any s with s ≧ 0
, Δs ≧ a (exp (1 / A) −1) and Δs is approximately a (exp (1 /
A) -1). Since the luminance quantization unit is Δs, the luminance quantization error is Δs / 2. Therefore, a (exp (1
If the constants a and A are set so that / A) -1) / 2 is about the accuracy of the brightness s itself, the quantization error of the brightness in the image memory is significantly smaller than the accuracy of the brightness itself. Since it does not happen, it becomes possible to use the memory efficiently.

For example, when the constants a and A are determined so that a (exp (1 / A) -1) / 2 is equal to 1/2 of the precision δ of the brightness itself,

Since a (exp (1 / A) -1) / 2 = δ / 2,

A = 1 / log (δ / a + 1) ≈a / δ Therefore, a / A can be determined as a / A = δ.

[Case of Edge Detection] Next, a case where edge detection is performed by a differential operator will be described. In the case of the present embodiment, the difference in density value between the pixel corresponding to the brightness s ₂ and the pixel corresponding to the brightness s ₁ is

L (s ₂ ) −L (s ₁ ) = A · log ((s ₂ + a) / (s ₁ + a)). In the case where s is sufficiently larger than a, that is, s >> a,

L (s ₂ ) −L (s ₁ ) ≈A · log (s ₂ / s ₁ ) Further, in the case where s is sufficiently smaller than a, that is, s << a,

[Equation 36] L (s ₂ ) −L (s ₁ ) ≈A / a · (s ₂ −s ₁ ). Since s ≧ 0, a> 0, and A> 0,

| L (s ₂ ) −L (s ₁ ) | <| A · log (s ₂ / s ₁ ) | | L (s ₂ ) −L (s ₁ ) | <| A / ( Since s ₂ −s ₁ ) |, it can be seen that the intermediate part between the two parts has properties in between.

When μ is an arbitrary number of μ ≧ 1, A> 0 and a> 0. Therefore, when 0 ≦ s ₁ ≦ s ₂ ,

If 0 ≦ A · log ((s ₂ + a) / (s ₁ + a)) ≦ A · log ((μs ₂ + a) / (μs ₁ + a)) s ₁ > s ₂ ≧ 0,

The relationship of 0 ≧ A · log ((s ₂ + a) / (s ₁ + a)) ≧ A · log ((μs ₂ + a) / (μs ₁ + a)) is mathematically established. Therefore, if s ₁ ≦ s ₂ ,

Where 0 ≦ L (s ₂ ) −L (s ₁ ) ≦ L (μs ₂ ) −L (μs ₁ ) s ₁ > s ₂ ,

Since the relationship of 0 ≧ L (s ₂ ) −L (s ₁ ) ≧ L (μs ₂ ) −L (μs ₁ ) holds, the difference value of the part where s ₁ and s ₂ is small is The weighting is smaller than the difference in the part where s ₁ and s ₂ are large. Therefore, the smaller the density value is, which is relatively inaccurate, the smaller the weighting of the difference value, and the above problem is solved.

It is possible to determine the value of the constant a based on the relationship between the accuracy of the brightness itself and the accuracy of the required brightness ratio. That is, the accuracy of the brightness itself is ± δ, and s ₂ >
Assuming s ₁ , the weighting of the difference value is

L (s ₂ + δ) −L (s ₁ −δ) = A · log (s ₂ / s ₁ ). That is,

It is necessary that {A · log (s ₂ + δ + a) + B} − {A · log (s ₁ −δ + a) + B} = A · log (s ₂ / s ₁ ). Solving the above equation for a,

(44) a = δ (s ₂ + s ₁ ) / (s ₂ −s ₁ ) = δ (s ₂ / s ₁ +1) / (s ₂ / s ₁ −1).

Therefore, if the required accuracy of the luminance ratio is γ, the constant a is

It can be determined that a = δ (γ + 1) / (γ−1). Also, according to the above-mentioned condition for efficiently using and storing the memory, a (exp (1 /
The constant A can be set so that A) -1) / 2 is approximately the accuracy degree δ of the brightness itself. For example, a (exp (1 /
When the constant A is set so that A) -1) / 2 is equal to 1/2 of the precision δ of the brightness itself,

Since a (exp (1 / A) -1) / 2 = δ / 2, the constant A is

It can be determined that A = 1 / log (δ / a + 1) ≈a / δ.

Further, the dynamic range of luminance is 0 ≦
When s ≦ smax, the maximum density value Lmax that must be able to be stored in the memory is Lmax = A ·
log (smax + a) −A · log (a). Therefore, the number of memory bits required for each pixel is log ₂ (Lmax
) Can be decided.

As an example, s ₁ = 1000, s ₂ = 15
00, s ₃ = 100, s ₄ = 150, s ₅ = 10, s ₆
= 15 and the accuracy of the brightness itself is ± 1, the difference value and the accuracy thereof are shown. In this case, s ₂ / s ₁
= S ₄ / s ₃ = s ₆ / s ₅ = 1.5, but it is necessary to distinguish this difference value from the difference values of other parts on the image, and the required accuracy of the luminance ratio Let γ be determined as γ = 1.1. Then, from the above equation, a = δ (γ +
Since 1) / (γ−1) = 21, the constant a is set to 21. Further, from the above equation, since A≈a / δ = 21, the constant A is set to 21. Then, as shown in FIG.

[0050]

L (s ₂ ) −L (s ₁ ) ≈8 ± 0.03 L (s ₄ ) −L (s ₃ ) ≈7 ± 0.3 L (s ₆ ) −L (s ₅ ) ≈ It becomes 3 ± 1, and it can be seen that the weighting decreases as the accuracy decreases.

However, the quantization error when storing in the image memory was ignored. The luminance quantization unit Δs when stored in the image memory is Δs = (s + a) as described above.
Since it is (exp (1 / A) -1), when the quantization error Δs / 2 of each luminance is calculated, s ₁ = 1000 ± 25, s ₂ =
1500 ± 37, s ₃ = 100 ± 3, s ₄ = 150 ±
4, s ₅ = 10 ± 0.8, s ₆ = 15 ± 0.9.
Therefore, when the effect of this quantization error is calculated and added to the accuracy of the difference,

[0052]

L (s ₂ ) −L (s ₁ ) ≈8 ± 0.03 ± 1 L (s ₄ ) −L (s ₃ ) ≈7 ± 0.3 ± 1 L (s ₆ ) −L ( s ₅ ) ≈3 ± 1 ± 1.

As can be seen from the above description, the first
In the second embodiment, the shape of the function suddenly changes from logarithmic to a proportional relationship at the boundary of s = a, whereas in the second embodiment, s = a.
In the vicinity of, the form of the function gradually changes from logarithmic to proportional relationship. Therefore, the second embodiment, compared to the first embodiment,
There is an advantage that the difference value is weighted according to the magnitude of the luminance over a wide range of the luminance.

In the above first and second embodiments, the above-mentioned predetermined function is used to convert the brightness into the density value.
It is obvious that the same effect can be obtained by using another function that approximates these functions instead of these functions, and these cases are also included in the present invention.

The conversion characteristics of the first and second embodiments are used.
Another embodiment of the edge detecting apparatus will be described in which the present invention is applied to an apparatus for capturing an image of a wide dynamic range and obtaining an edge image from the image. The configuration of this device is shown in FIG. FIG. 9 shows the overall flow of processing performed by this device.

This apparatus operates as follows and picks up an image with a wide dynamic range. First, the exposure time control device 13 sets the exposure time (shutter speed) of the imaging device 12 to 1/60 seconds (step 100, hereinafter,
An image is taken (S102). Note that the γ characteristic of the image pickup device 12 is assumed to be adjusted to γ = 1. The picked-up image data is read from the image pickup device 12 for each pixel, converted into digital data by the A / D converter 14, transferred to the image memory (1) 21 through the image data I / O 15, and stored therein. (S104).
Next, the exposure time control device 13 sets the exposure time of the imaging device 12 to 1/300 seconds (S106), and imaging is performed (S108). The picked-up image data is read from the image pickup device 12 for each pixel, converted into digital data by the A / D converter 14, transferred to the image memory (2) 22 through the image data I / O 15, and stored therein. (S11
0).

In general, the image pickup device 12 has the same brightness.
The signal intensity of the image data output by is proportional to the exposure time unless the output of the imaging device 12 is saturated. Therefore, the image memory (1) 21 and the image memory (2) 22
The relationship between the density value L and the brightness s of the image is as shown in FIG. Next, in S112, as shown in FIG.
The CPU 10 calculates a wide dynamic range image from the data in the image memory (1) 21 and the image memory (2) 22, and stores the wide dynamic range image in the image memory (3) 23.

In FIG. 10, L ₁ (x, y), L
₂ (x, y) and L ₃ (x, y) represent the density values of the pixel (x, y) of the image memory (1) 21, the image memory (2) 22 and the image memory (3) 23, respectively. Also, X
max and Ymax represent the number of pixels in the x direction and the number of pixels in the y direction of the image, respectively. Further, L _m represents the output value of the A / D converter 14 corresponding to the maximum image signal level at which the output of the image pickup device 12 is not saturated. Also, F (s)
Represents a function for converting the brightness s into the density value of the image memory (3) 23.

As described above, the region 1 (0≤s≤s in FIG.
The data of the luminance range of ₂ ) is the data of the image memory (1) 21, and the data of the luminance range of the area 2 (s ₂ ≦ s ≦ s ₃ ) is the data of the image memory (2) 22. Image data of both luminance areas (0 ≦ s ≦ s ₃ ) of the area 2 are combined and stored in the image memory (3) 23. As can be easily understood, the data of the image memory (2) 22 needs to be five times as large as the data of the image memory (1) 21 in order to match the scale of the luminance at the time of composition.

A method for obtaining an image having a wide dynamic range in the image memory (3) 23 in the format obtained by converting the first embodiment and the second embodiment described above will be described here.

[When logarithmic conversion is performed over the entire area] An image having a wide dynamic range is converted using the function A · log (s) + B so that the density value thereof is proportional to the logarithm of the brightness over the entire brightness region, and the image memory is converted. (3) Consider the case of storing in 23. The luminance quantization unit Δs is, as described above,

[Expression 50] Δs = s (exp (1 / A) −1) ≈s / A (A >> 1)

On the other hand, when the number of bits of the A / D converter 14 is set so that the quantization accuracy in the A / D converter 14 is about the noise level of the image data output from the image pickup device 12, the brightness itself is determined. The accuracy δ is considered to be approximately the quantization accuracy in the A / D converter 14. In the combined wide dynamic range image, the quantization accuracy in the A / D converter 14 becomes 5 times that of the exposure time in the area 2 compared to the area 1. Therefore, the accuracy of the luminance itself is δ in the area 1 and 5δ in the area 2.

By the way, the maximum value Lmax of the density value in the image memory (3) 23 is as follows when the maximum value of the brightness is smax.

[Expression 51] Lmax = A · log (smax) + B. Therefore, the larger the constant A is, the more the density value of the image is increased, so that the image information stored in the image memory (3) 23 is increased. Therefore, from this viewpoint, the constant A should be as large as possible.

On the other hand, as described above, if Δs is smaller than the accuracy of the luminance itself, the density value of the memory is redundantly used, which is not preferable. From the above, for example, the accuracy δ of the brightness itself and the quantization unit Δ of the brightness
If the constant A is determined so that the relationship with s becomes as shown in FIG. 13, Δs is obtained in the low luminance part (0 ≦ s ≦ s ₁ ).
Is smaller than the accuracy of the brightness itself, and the density value of the memory is used redundantly. In order to avoid this problem, it is conceivable to determine the constant A so that the relationship between the accuracy of the brightness itself and the quantization unit of the brightness is as shown in FIG. 14, but in this case the slope of Δs However, since the value of A becomes smaller due to the increase of the number of pixels, there is a problem that the number of density values of the image becomes smaller. That is, the former has a large number of density values of the image but has a large redundancy of the density value, and the latter has a problem that the redundancy of the density value is small but the number of the density values of the image is small and the information of the image is small. ..

[When the Conversion Characteristics of the First and Second Embodiments are Used] Next, the present invention is applied to the present apparatus, and an image having a wide dynamic range is obtained with the density value L and the brightness s of the image. Of the function L = A · log (s shown in the second embodiment described above.
A case will be described in which the image data is converted into the relationship of + a) + B and stored in the image memory (3) 23. In this case, the luminance quantization unit Δs is, as described above,

[Expression 52] Δs = (s + a) (exp (1 / A) −1) ≈ (s + a) / A (A >> 1)

Therefore, for example, if the relationship between the accuracy of the brightness itself and the brightness quantization unit is as shown in FIG.
Since Δs is always larger than the accuracy of the brightness itself, the density value of the memory is not used redundantly, and the inclination of Δs is small, the value of A is large, so that the number of density values of the image is large. That is, when the present invention is used, an image having a wide dynamic range in which the number of density values of the image is large and the redundancy of the density value is small is obtained in the image memory (3) 23 as compared with the above-mentioned conventional technique. It will be possible. When determining the constant A as shown in the figure, as can be seen from the figure,

Δ = a / A 2 δ = s ₁ / A + a / A. Therefore, the constant a is

[Expression 54] a = s ₁ .

The same effect can be obtained when the function shown in the first embodiment is used. In this case, the luminance quantization unit Δs is, as described above,

Δs = s (exp (1 / A) −1) ≈s / A (s ≧ a, A >> 1) Δs = a / A (s <a)

Therefore, if the relationship between the accuracy of the brightness itself and the quantization unit of the brightness is shown in FIG. 16, Δs is always larger than the accuracy of the brightness itself, and the density value of the memory is redundantly used. Since the value of A is large because the gradient of Δs is small, the number of density values of the image is large. That is, when the present invention is used, an image having a wide dynamic range in which the number of density values of the image is large and the redundancy of the density value is small is obtained in the image memory (3) 23 as compared with the above-mentioned conventional technique. It will be possible. In this case, as can be seen from the figure, the constant a is defined as a = s ₁ .

Next, in this apparatus, the image memory (3)
A processing procedure for obtaining an edge image from the wide dynamic range image stored in 23 will be described. It is executed according to the program shown in FIG. 11 which is processed in S114 of FIG. The image stored in the image memory (3) 23 is
By applying the bel operator and storing the result in the image memory (4) 24, the image memory (4) 2
An edge image can be obtained at 4. Image memory (4)
The edge image of 24 is an image BU as shown in FIG.
D / A converter 17 from S through image data I / O 16
Can be converted into an analog video signal and viewed on the video monitor 18. In addition, in FIG. 11, abst (M) is
Represents the absolute value of M. Further, L ₃ (x, y) is the density value of the pixel (x, y) of the image memory (3) 23 and is L ₄ (x, y).
Represents the density value of the pixel (x, y) of the image memory (4) 24. Further, Xmax and Ymax represent the number of pixels in the x direction and the number of pixels in the y direction of the image, respectively. As described above, when the brightness is converted by using the present invention and the image is stored in the image memory (3) 23, as described above, a portion having a relatively low accuracy and a small density value is used. As the difference value is weighted more, the conventional problem is solved.

[Brief description of drawings]

FIG. 1 is a characteristic diagram showing a relationship between luminance and a density value which is a conversion value when logarithmic conversion is performed on all areas.

FIG. 2 is a characteristic diagram showing conversion characteristics from luminance to density values in the device according to the first embodiment of the present invention.

FIG. 3 is a diagram showing a result of difference calculation of density values obtained by logarithmically converting luminance in all regions, showing characteristics regarding a central value of difference calculation and an error in a luminance region at the same luminance ratio. Characteristic diagram.

FIG. 4 shows results of difference calculation of density values when the conversion characteristic shown in FIG. 2 is used in the device according to the first embodiment, and shows the difference between the center value and the error of the difference calculation at the same luminance ratio. FIG. 6 is a characteristic diagram showing characteristics relating to a luminance region.

FIG. 5 is a characteristic diagram showing the conversion characteristic from the luminance to the density value in the device according to the second embodiment of the present invention, together with the approximation characteristic in the low luminance region and the high luminance region with the luminance being a logarithmic scale.

FIG. 6 is a characteristic diagram showing, on a linear scale, conversion characteristics from luminance to density values in the device according to the second embodiment of the present invention.

FIG. 7 shows the result of difference calculation of density values when the conversion characteristic shown in FIG. 6 is used in the device according to the second embodiment, and shows the central value and error of difference calculation at the same luminance ratio. FIG. 6 is a characteristic diagram showing characteristics relating to a luminance region.

FIG. 8 is a block diagram showing a configuration of an image conversion apparatus according to first and second embodiments.

FIG. 9 is a flowchart showing a procedure of data processing by a CPU in the embodiment apparatus.

10 is a flowchart showing a procedure for synthesizing a wide dynamic range image in the data processing procedure shown in FIG.

11 is a flowchart showing an edge image generation processing procedure of the data processing procedure shown in FIG.

FIG. 12 is a characteristic diagram showing the relationship between the ratio of the error of the output signal obtained from the image pickup device and the luminance itself to the quantization error of the luminance and the extended luminance range.

FIG. 13 is a characteristic diagram showing the relationship between the luminance quantization error and the luminance when the luminance is logarithmically converted for all areas together with the precision of the luminance itself.

FIG. 14 is a characteristic diagram showing a method of increasing the quantization error of luminance in the low luminance region to be higher than the precision of the luminance itself.

FIG. 15 is a characteristic diagram showing a method according to a second embodiment in which a luminance quantization error is made larger than the accuracy of luminance itself in a low luminance region.

FIG. 16 is a characteristic diagram showing a method according to the first embodiment in which a luminance quantization error is made larger than the accuracy of luminance itself in a low luminance region.

[Explanation of symbols]

 10 ... CPU (converting means) 23 ... Image memory Step 112 (step of FIG. 10) ... Converting means

Claims (1)

[Claims] 1. In an image conversion device for converting the brightness of an image into a density value based on a predetermined characteristic and creating an image represented by the density value in an image memory, in an area where the brightness of the image is relatively high, The conversion means converts the brightness into a density value by a characteristic substantially proportional to the logarithm of the brightness, and in a region where the brightness of the image is relatively small, the conversion means converts the brightness into a density value by the characteristic substantially proportional to the brightness. An image conversion apparatus, wherein an image is created by storing density values output by the image memory in the image memory.