JP2004007826A - Picture correcting method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To obtain correct picture data in a digital picture reading device to be used for information equipment and an OA apparatus, etc. <P>SOLUTION: Errors on a temporal axis and a frequency axis are obtained by reading a test chart. The first error as the one on the temporal axis, a first error main cause as the main cause of the first error, the second error as the one on the frequency axis, and a second error main cause as the main cause of the second error are respectively and previously stored in a first database. Thereafter an original and a picture are read, from this reading data are obtained on the temporal axis and frequency axis, a corresponding relation between the first error and the first error main cause and a corresponding relation between the second error and the second error main cause are obtained from the above reading. Then the corresponding first and second error main causes are picked-up, and a correction algorithm is performed to cancel the errors in picture data based on the picked-up first and second error main causes. <P>COPYRIGHT: (C)2004,JPO

Description

The present invention relates to an image correction method.

で は Information devices, OA devices, and the like frequently perform work of reading images and outputting them to displays, paper sheets, and the like. For this purpose, an image reading device is used. FIG. 30 shows a block diagram of a conventional scanning image reading apparatus. In FIG. 1, a document 2 is placed on a document table 3 made of glass, a cover, or the like. The contact type reading head 1 moves in parallel in the direction of arrow 6 in FIG. This parallel movement is performed by forward rotation and reverse rotation of the motor 8, the pulley 9, and the belt 10. The reading head 1 includes a lamp 15 for irradiating a document, a CCD 4 for converting a light image into an electric signal, and an optical system 5 for forming pixels of the document on the CCD. The output of the CCD 4 is converted into a digital value corresponding to the image density by the A / D converter 7 and stored in the storage unit 11 via the interface control unit 12. Digital values carrying image density information are spatially discrete by A / D operations at predetermined time intervals. The interface control unit 12 supplies a clock signal to the A / D converter to synchronize the capture of the digital value and the movement of the read head, and controls the operation of the motor drive unit 14 via the motor control unit 13. ing. The motor drive unit 14 supplies electric power to the motor 8 according to a drive signal from the motor control unit 13. The image data stored in the storage unit 11 is taken in, for example, a computer of an electronic publishing system via an external data bus (not shown), and is used as a picture in a printed document.

In the image reading device having such a configuration, for example, when the rotation unevenness of the motor 8, the tracking of the belt 10, and the dirt on a guide rail (not shown) for guiding the movement of the reading head 1 occur, the moving speed of the reading head 1 fluctuates. At the time of reading an image, the image is relatively expanded and contracted. Further, when each mechanism constituting the apparatus resonates due to, for example, vibration of the motor 8, a reading position error in the reading head 1 may occur, which is a factor of image data error.

Therefore, in order to reduce the vibration in the mechanical structure for scanning the read head 1, the mechanism is generally made to have high rigidity, so that the design specification becomes severe and the apparatus cost increases. In some cases, the adjuster can reduce the reading error by adjusting the apparatus. However, image data such as uneven rotation of the motor 8, insufficient rigidity of parts of the apparatus, poor connection, occurrence of an excessive assembly error, and poor adjustment of the CCD 4. There are many factors that affect the error of the adjustment, and the adjustment requires a long time and the coordinator's experience because they influence each other.

(4) As described above, in the image reading apparatus that captures image information while the reading head moves, there is a problem that if the moving speed and the moving position of the reading head fluctuate, the obtained image data is different from the true one. For this reason, it is necessary to set the image data to a value as close to the true value as possible by adjustment by the adjuster, correction of the image data, and the like. However, there are many error factors in the image data, and there is a correlation between them, so that there is no appropriate method for determining the adjustment method and the image data correction method.

Accordingly, it is an object of the present invention to provide a method capable of obtaining image data in which errors are reduced as much as possible from image data containing errors.

In order to achieve the above object, the present invention obtains an error on a time axis and a frequency axis by reading a test chart, and obtains a first error as an error on a time axis and a first error factor as a factor thereof. And a second error as an error in the frequency axis and a second error factor as a factor thereof are stored in the first database in advance, respectively.
Thereafter, an image of the target document is read, and the first error and the first error factor in the first database are determined based on data on the time axis and the frequency axis obtained from the read image data. And the corresponding first and second error factors are picked up from the corresponding relationship between the second error factor and the second error factor,
And executing a correction algorithm for canceling an error occurring in the image data based on the first and second error factors picked up.
It is characterized by the following.

According to the present invention, the read image data can be properly corrected.

FIG. 1 is a functional block diagram showing a configuration of an apparatus relating to an embodiment of the present invention. The image reading mechanism 51 includes a reading head, a scanning mechanism, and the like, and a light image formed by a CCD or the like is converted into an electric signal. An image reading unit 52 for obtaining data, a mode selecting unit 53 for selecting a flow of image data according to the setting of an operation mode of the image reading apparatus, an image data error feature extracting circuit 54 for extracting an error feature from the image data, An error factor estimating unit 55 for estimating a factor causing an error, an adjusting method determining circuit 56 for determining an adjusting method from the estimated error factor, an image correcting unit 57 for correcting image data by a set correcting method, A storage unit 58 for storing the read image data, a horizontal position adjusting mechanism for the reading head in the main scanning direction, A mechanism control unit 59 that controls the operation of various adjustment mechanisms 63 such as a position adjustment mechanism, a height adjustment, a guide rail balance adjustment mechanism, a switching mechanism of the rail to a new slide surface, and a drive belt tension adjustment mechanism is estimated. Based on the error factors, adjustment of the drive torque to the drive motor for driving the read head, adjustment of the rotation speed, control constant of a feedback control loop for moving the read head at a constant speed at a specified target speed (not shown), and the like are performed. The motor control unit 60 includes a motor drive circuit 61 that power-amplifies the output of the motor control unit 60 and drives the motor. The mode selector 53 to the motor controller 60 are formed on a microprocessor board. The mode selection unit 53 to the adjustment method determination unit 56 constitute the interface 12.

In this configuration, when the normal mode is selected by the mode selection unit 53, the operation is performed in the same manner as the conventional image reading apparatus described above, and the image data obtained by the image reading unit 52 is stored in the storage unit 58 as it is. At this time, if the error adjustment described later has already been executed and the image correction by the predetermined method has been set, the image data corrected through the data correction processing by the image correction unit 57 is stored in the storage unit 58. Also, even when the mechanism control corresponding to the image data error has already been performed, the motor driving and the like are performed by the control method.

When the image adjustment mode is selected in the mode selection unit 53, the image data error feature extraction unit 54, the error factor estimation unit 55, and the adjustment method determination unit 56 are used to execute various adjustments including mechanism control, electrical control, image correction, and the like ( 57, 59, 60, 61 and 63), an appropriate adjustment is made to reduce the image data error. In the normal mode, the mechanism control and image correction determined at this time are used. When manual adjustment is required, a predetermined adjustment position and adjustment method are displayed on the display of the information display unit 62. Hereinafter, data processing in the above-described main part will be described.

(Image data error feature extraction unit)
FIG. 2 is a block diagram showing a data processing procedure in the image data error feature extraction of the present invention, and each processing will be described.

(1) Image data creation
First, creation of image data used for testing an apparatus will be described. FIG. 3 shows an example of a test chart used to extract the feature of the error from the image data, in which a large number of parallel lines of equal pitch are printed. When the reading head scans the test chart placed on the document table at a constant speed in the Y-coordinate direction in the drawing while reading a pixel row in the main scanning direction in the X-coordinate direction, a certain X Y coordinate direction of the concentration distribution of the coordinate, i.e. the data sequence d of the image data in the sub-scanning direction _{(x i, y 0),} d (x i, y 1), d (x i, y 2), ... is The distribution is such that a sine wave as shown in FIG. 4 is spatially discretized. This is time-series data sampled at a certain fixed time and discretized over time. In this example, the density data is represented by 8 bits. Therefore, image data is obtained in the density level range of 0 to 255. This data sequence is used as image data used for extracting the image data error feature in the present invention.

Normally, the length of the test image data covers the entire range in which the read head moves, but if it is known that the data varies greatly at a particular location, only the data corresponding to that location is targeted. The following processing may be performed. In this case, there is an advantage that the data processing time can be shortened because the number of data is reduced.
Further, by using the data sequence d (x ₀ , y _i ), d (x ₁ , y _i ), d (x ₂ , y _i ),... In the main scanning direction, the error factor for the data in the main scanning direction is specified. Is effective for The coordinator can arbitrarily select a data string.

(2) Data normalization
By the way, in the example shown in FIG. 4, the width of the maximum value and the minimum value of the data is much smaller than the density level range of 0 to 255. As described above, even if the A / D converter that converts the read electrical signal into image data is 8 bits, the level change range of the data actually output from the A / D converter is not necessarily in the range of 0 to 255. The range of the maximum value and the minimum value of the data obtained varies depending on the sensitivity of the CCD used, the light intensity of the optical system, the density and resolution of the test chart used, and the like. Therefore, the data is normalized as shown in FIG. 5 so that the data obtained under any conditions can be handled in the same manner in the subsequent data processing method.

The following equation (1) is used for data normalization, where the maximum value d _max and the minimum value d _min of the target data string are set.
^{_{d * (x i, y i}} )
_{= Do × (d (x i} , x i) -d min) / (d max -d min) ... (1)
Here, d ^* is normalized data, and Do is a decimal representation of the number of bits of the A / D converter.

(3) Data interpolation
The image data is spatially discretized. For example, if the sampling time is ΔT [sec] and the moving speed of the read head is V [mm / sec], the length of ΔY = ΔT × V [mm] Data is sampled for each (quantized reading range). Therefore, due to restrictions on the performance of the circuit for capturing image data, the driving mechanism of the optical system, and the like, if the sampling time becomes longer or the moving speed of the reading head becomes slower, the reading range ΔY per scanning becomes longer. For this reason, the analysis accuracy of the reading position error decreases. Therefore, data interpolation is performed to obtain data at a position finer than the actual reading range ΔY. As a data interpolation method, any method such as so-called linear approximation, regression curve approximation, Lagrange interpolation, and spline interpolation may be used.

Here, a data interpolation method using mixed spline interpolation among the spline interpolations will be described. The mixed sfline interpolation is performed by using an interpolation formula such as the following equation (2) with four consecutive points as one set, and using only the central two points. u is an interpolation parameter, and has a range of 0 ≦ u ≦ 1.
_{d (x i, y i +} u)
= -U × (u-1) × (u-2) / 6 × d (x i, y 0)
+ (U + 1) × (u−1) × (u−2) / 2 × d (x _i , y ₁ )
− (U + 1) × u × (u−2) / 2 × d (x _i , y ₂ )
+ (U + 1) × u × (u−1) / 6 × d (x _i , y ₃ ) (2)
(2) gradually increases by 0.1 to u using equation when _{d (x i, y 1)} , d (x i, y 2) so that the interpolated 10 between.

FIG. 6 shows an example in which mixed spline interpolation is performed. For example, in the case of an image reading apparatus having a resolution of 600 [dpi, dot per inch], ΔY is 0.042 [mm], and this interpolation results in image data of every 0.0042 [mm]. .
This interpolation is equivalent to performing a simple data restoration between the discretized data. However, not all data can be restored. According to the results of the study by the inventor, in the case of an image reading apparatus having a resolution of N [dpi], the effect of the present invention is easily obtained when a test chart having a resolution of N / 3 or less is used. Similarly, in the case of an image reading apparatus in which the moving speed of the read head is made variable to perform image zooming, the sampling time is used as ΔT [sec], and the read head moving speed is used as V [mm / sec]. It is effective to use a test chart smaller than the resolution obtained by the equation (3).
N _test = 1/3 / V / Δt (3)

(4) Reading position error calculation
The reading position error, obtaining the pitch P _i of the peaks and valleys from the data string level value becomes a sine wave as shown in FIG. Subsequently, the the P _i, the pitch _{P i-1, P i,} P i + 1, P i + 2, ... average pitch Pm or, position error data string by using the ideal pitch Po where no positional deviation of the Δi is obtained from the following equation.
Δi = Pi−Pm (or Po) (4)
FIG. 8 shows an example of an error distribution in the Y direction of the position error Δi when the test chart is scanned, that is, in the sub-scanning direction. The scanning position in the sub-scanning direction is indicated by the line number of a grid-like parallel line printed on the test chart. Since this line number corresponds to the position on the time axis in scanning, the graph shown in FIG. 8 represents so-called jitter, which is the fluctuation of the reading speed on the time axis due to the reading position error Δi. This Δi can be used as an evaluation value (feature parameter) for the reading position error. That is, when the amplitude of Δi is large, the image data is distorted as compared with the image of the document. As evaluation values (feature parameters) for the entire image, J parameters such as a maximum value Jmax, a minimum value Jmin, an average value Jave, a variance Jψ, a standard deviation Jσ, and the like are used.

If △ i is locally large, the △ i data string corresponds to the sub-scanning position, so that an abnormal portion can be estimated. For example, when the peak of △ i repeatedly occurs at the same position, dirt / scratch of the guide rail corresponds thereto, and when it occurs at a constant cycle, tracking in the drive system corresponds thereto. If the amplitude value distribution has a slope, the inclination of the guide rail or the like corresponds to the slope. When there are a plurality of defective portions, they usually appear in a composite form in the graph of △ i. The correspondence between the state of various characteristic parameters serving as a reference for estimating the cause of the error and the cause of the error is stored in a ROM (not shown) in a database. In the main scanning direction in the X direction, the evaluation of the apparatus performance and the cause of the error can be similarly performed.

(5) Spectrum distribution calculation
Since the reading position error Δi is also time-series data, it can be considered that the reading position error Δi is equivalent to the speed fluctuation (jitter) of the reading head at each sampling time interval. Therefore, when frequency analysis is performed using the data sampling time, frequency components for each frequency can be extracted. As a method of extracting the frequency component, any method such as a so-called fast Fourier transform (FFT) and a maximum entropy method (MEM) may be used.

FIG. 9 shows an example in which the spectrum distribution of △ i is obtained by FFT. Peak S ₁ (28 Hz, level a), peak S ₂ (56 Hz, level b), peak S ₃ (76 Hz, level c), and peak S ₄ (84 Hz, level d) appearing in this graph are characteristic parameters. , Respectively, responds to vibration inherent in each mechanism of the image reading apparatus, rotation fluctuation of the driving motor, and the like. When the vibration of each mechanism or the rotation fluctuation of the driving motor is small, the corresponding peak disappears or the peak value decreases. When the image is blurred as a whole, the bandwidth of the peak is widened. Therefore, the correspondence between the frequency f and peak value p of the peak S, the mechanism that generates this peak, and the corresponding relationship is stored in the ROM in advance. Then, referring to the database, the S-parameter (f, p) based on the measured peak frequency f and peak value p determines which mechanism causes a vibration that has an adverse effect on the corresponding reading operation in a later circuit. It can be determined.

(Error factor estimator)
FIG. 10 is a block diagram illustrating a flow of processing in the error factor estimating unit 55 according to the present invention. Here, the evaluation of how accurately the image of the original was read by the image reading apparatus using the J parameter relating to the reading position error and the S parameter relating to the spectral distribution obtained by the image data error characteristic extracting unit 54, and the error factors are as follows. Do something analysis.

{Circle around (1)} First, the J parameter (Jmax, Jmin, Jave,...), Which is the image error evaluation value due to the reading position error, is compared with their allowable errors Jo (Jmaxo, Jmino, Javeo,...). Further, the S parameter is compared with the S parameter So corresponding to the allowable image quality (55a in FIG. 10). Data relating to the allowable error Jo and the allowable image quality So is stored in the database 55e in advance. As a result of the comparison, when the values of the J parameter and the S parameter of the read image data are smaller than the permissible error Jo and the permissible image quality So, there is no need for adjustment, and the determination of the error factor ends (flag K = 1).

When the values of the J and S parameters are larger than the permissible reference values Jo and So, that is, when both or one of the permissible error and the image quality in the image data is not permissible, the first estimator 55b sets the J parameter to Estimation processing based on this is performed. A large number of error factors corresponding to combinations of the values of each J parameter are stored in the database 55e in advance, and errors corresponding to the extracted values of each J parameter or the degree of correspondingness exceeding a predetermined discrimination value are stored. Pick up the factor from the database. Further, as described above, when the reading position error 述 i is locally large, the △ i data string corresponds to the sub-scanning position, and the factor corresponding to the error occurrence position is fetched from the database 55e. When the peak of Δi repeatedly occurs at the same position, for example, dirt / scratch of the guide rail corresponds. When the peak occurs at a constant period and the position of occurrence changes, tracking in the drive system corresponds to the same. If the distribution of the amplitude value of Δi has a slope, the slope of the guide rail or the like corresponds. In the case where there are a plurality of defective portions, they usually appear in the graph of △ i, and thus in the J parameter.

Next, an error factor is estimated based on the S parameter using the spectrum distribution. The spectrum S parameter corresponds to the inherent vibration of the rotating component constituting the image reading mechanism, the rotation fluctuation of the driving motor, and the like. Therefore, it is possible to search for the natural frequency of the component previously stored in the database 55e by using the S parameter and to estimate the corresponding component. For such estimation, fuzzy inference, neural network, expert system, statistical processing, etc. can be used as appropriate (55c).

(4) Comprehensive judgment is performed based on the error factors extracted by the first estimating unit 55b and the second estimating unit 55c. Some error factors can be more clearly discriminated by looking at both the J parameter, which is the error evaluation value of the jitter on the time axis, and the S parameter, which is the error evaluation value on the frequency axis. One error factor may appear in both the J parameter and the S parameter. Accordingly, each item of the J parameter and the S parameter is made to correspond to each node of the input layer, and the neural network using the parameter value as an input value to the node is used. If the degree of applicability exceeds a reference value determined for each error factor, it is determined that the error factor has occurred in the apparatus. Reference data necessary for the comprehensive estimation is stored in the database at 55e. It is preferable that the neural network performs learning using various parameter data whose error factors are known in advance, and tunes weight coefficients in the input layer, the intermediate layer, and the output layer by, for example, back propagation learning.

As described above, the error factor determined by the J parameter which is the error evaluation value on the time axis, the error factor determined by the S parameter which is the error evaluation value on the frequency axis, and both the J parameter and the S parameter are determined. Obtain information about error factors.

{Circle around (5)} The overall estimating unit 55d determines the image quality. Here, the normal spectral distribution of the reference (test) image already provided to the database 55e is compared with the input image data, and the state of image quality is estimated based on the density, resolution, vibration components, and the like. For the estimation of the image quality, an appropriate method such as fuzzy inference, a neural network, an expert system, and statistical processing can be used in addition to the method of comparing with the normal data as described above.

In this embodiment, the case where the reading position error and the spectrum distribution are used as the image error evaluation value is described. However, the image error of the apparatus such as the distribution of the image data itself and the spectrum can be evaluated, and the correlation with the error factor can be evaluated. Any quantity can be applied.

Further, if the obtained spectrum is further separated into waveforms and the peak value for each frequency is accurately determined, each error factor can be specified more accurately. The waveform separation method may be any of a simplex method, a DFP method, a GN method, and the like.

As the image error evaluation value due to the reading position error, a maximum value, a minimum value, an average value, a standard deviation value, or the like may be used instead of using the distribution as it is. However, since the reading error has both positive and negative values, the absolute value may be averaged in some cases. In the case of a spectral distribution, the data is used in the form of a pair of a frequency and a peak value appearing (p _i , f _i ), or a gain data string (p ₁ , p ₂ ,..., P _i , …) Is used. Further, when a spectrum distribution is obtained by a complex number like a Fourier transform, a form of a pair (R _i , I _i ) of a real part and an imaginary part may be used.

(Adjustment method determination unit)
As described above, when the read image data is smaller than the allowable error and the image quality is allowable (flag K = 1), the image data is stored in the storage unit 58 via the image correction unit 57. If either or both of the allowable error and the image quality cannot be tolerated, the error factor estimating unit 55 estimates the error factor, and outputs an error factor code representing the estimated error factor.

The adjustment method determining unit circuit 56 includes, for example, an adjustment content determining unit 56a, an image correction content determining unit 56b, and a database 56c shown in FIG. The adjustment content determining unit 56a reads a corresponding processing algorithm from the database 56c based on the error factor code supplied from the error factor estimating unit 55. This processing algorithm is roughly classified into an adjustment algorithm for motor control (flag K = 2), a mechanism adjustment algorithm for adjusting the image reading mechanism (flag K = 3), an algorithm for correcting image data (flag K = 4), When there is no mechanism for automatically adjusting the error factor, or when the error factor cannot be specified, it is classified into a manual adjustment algorithm (flag K = 5) that assists the assembly coordinator of the device to adjust manually. Is done. The processing algorithm may include a plurality of algorithms described above.

(4) Information for reducing the error factor, such as the constant of the feedback speed control loop and the amplitude and phase of the motor drive pulse train, are sent to the motor control unit 60 by the motor control adjustment algorithm. The mechanism adjustment algorithm instructs the mechanism controller 59 to increase or decrease the set value of the corresponding adjustment mechanism. For example, a read head balance adjustment mechanism, a drive belt tension adjustment mechanism, a guide rail balance adjustment mechanism, and the like are adjusted. A correction algorithm for canceling an error generated in the image data due to a specific error factor is set in the image correction unit 57 by the algorithm for correcting the image data, and the image data is corrected.

When the image adjustment mode is selected in the mode selection unit 53, the adjustment method in each adjustment circuit at that time is set. When the image adjustment mode is selected again, a new adjustment method is updated according to the state of the image data at that time. Therefore, even if aging occurs and previous adjustments are no longer appropriate, selecting the image adjustment mode allows the apparatus to be in an optimal state at that time.

Hereinafter, various adjustments in the adjustment method determining unit according to the present invention will be described.

(1) Motor control unit (flag K = 2)
The case where the distribution of the reading position error is shifted to positive or negative as shown in FIG. 12 is a case where the set value of the moving speed of the reading head is shifted. When the spectral peak of the distribution of the position error has a frequency related to the motor as shown in FIG. 13, there is a problem in driving the motor. The adjustment method determination unit 56 extracts an adjustment algorithm for motor control according to the determination result based on the correlation in the error factor estimation unit 55, and causes the motor control unit 60 to adjust the motor control parameters.

場合 In the case shown in FIG. 13, the target speed is updated to a predetermined speed. Further, when using the closed loop servo, there is a case where a steady-state deviation occurs with respect to the target speed. In such a case, adjustment is made by changing the gain of the servo circuit. For example, in the case of a digital servo using a DSP (Digital Signal Processor) or the like, automatic adjustment can be easily performed (flag K = 2). Further, when gain adjustment by the adjuster is necessary in the analog servo circuit, such an instruction or adjustment procedure is displayed on a display device 62 described later, and the adjustment is supported (flag K = 5).

The peak of the reading error as shown in FIG. 13 appears when the frequency fluctuates at a constant frequency around the target speed. In the closed loop servo, the constant of the control loop used in such a case is changed so as to have low vibration, or the sampling time is set sufficiently short with respect to the fluctuation frequency in which the problem occurs.
Further, the adjuster may attach an attenuator to the rotating shaft of the motor, and the display device 62 indicates the fact. In this case, an adjustment circuit for manual adjustment (K = 4) is also selected at the same time.

As shown in FIG. 14, the distribution of the reading position error sometimes having a large value sometimes appears. Such a distribution appears, for example, when the load is locally changed in a range in which the read head moves, and the vibration cannot be suppressed by the initially set control loop specifications. Even in such a case, a control loop that is sufficiently stable at a problematic frequency and has a high gain is configured.

The cause of the load fluctuation may be scratches, dirt, poor lubrication, etc. on sliding members such as guide rails. Therefore, there is a great possibility that the problem as shown in FIG. 14 can be solved by cleaning the portion or applying a lubricant. Also in this case, the adjustment circuit of the manual adjustment (K = 5) is simultaneously selected, and the adjustment method (cleaning and lubrication of the rail) is displayed on the display device 62.

It is to be noted that optimal regulator control (lqr or lqi), _H∞ control and the like are effective for such disturbance. When such control is used for the motor drive control, the weight of the evaluation function of the optimal regulator control and the frequency weight distribution of the _H∞ control are changed to calculate each constant of the control loop that is strong against disturbance. , For the subsequent motor control.

(2) Mechanism control unit (flag K = 3)
The mechanism control unit 59 controls the adjustment mechanism 63 related to the reading operation and the image quality of the image reading device 51 according to the command code given from the adjustment content determination unit 56a of the adjustment method determination unit 56 as described above. As the adjusting mechanism 63, there are various types as described above. For example, the reading head 1 has a balance adjustment mechanism for adjusting the height and parallelism of the CCD sensor, a slit adjustment mechanism for adjusting the light amount balance value of the light source, a lens position adjustment mechanism for adjusting the image distortion and the focal position, and optics. There is a cleaning mechanism for the system. The scanning system includes a guide rail cleaning mechanism, a rail pipe sliding surface shifting mechanism, a guide rail balance adjusting mechanism, a drive belt tension adjusting mechanism, and a variable damper mechanism for attenuating vibration of the rotating system. Various types of such adjustment mechanisms are provided according to error factors. By repeating the mechanism adjustment and the error extraction / estimation, it converges to a constant state allowed as the device specification, and the error factor is removed.

(3) Image correction unit (flag K = 4)
As image correction, edge enhancement, smoothing processing, and the like have been conventionally performed. The purpose of the image correction of this embodiment is to solve the mechanical and control problems of the apparatus, and it is not limited to such image correction.
The image correction method executed in the present invention may be any correction method that can achieve the object. In this embodiment, the image correction method using a neural network, the density correction method by estimating the charge amount of the CCD, Will be described.

An image correction method using a neural network When a test chart in which a large number of parallel lines are printed at the same pitch as shown in FIG. 3 is read in the Y coordinate direction in FIG. When the contact read head moves at a constant speed, the density data has a distribution in which a sine wave as shown in FIG. 15 is spatially discretized as described above. This is a value sampled at a certain fixed time and a value discretized in time. On the other hand, if there is a change in the moving speed of the read head, the time axis expands and contracts, and in many cases, an effect such as FM modulation occurs, resulting in a value different from the true value as shown in FIG. Such an error appears as a peak at a certain frequency when the reading position error is Fourier-transformed. The direct cause is vibration of the mechanical unit or speed fluctuation of the drive motor, and the corresponding adjustment circuit is also selected as an adjustment target at the same time.

The image correction mode is performed for reading position errors that can be corrected empirically. Therefore, the error factor estimating unit 55 confirms that the image correction range is already known according to the degree of the reading position error, and selects the image correction mode (flag K = 4).

FIG. 17 schematically shows the structure of the neural network used in the embodiment. In this example, there are three layers, an input layer, an intermediate layer, and an output layer, and each layer is composed of a multi-input, multi-output arithmetic element called a unit. The output X _{j of a} certain unit is expressed by the following equation (5). Here, f (x) is an arbitrary function, but a function having an S-shaped characteristic such as a sigmoid function represented by Expression (6) is usually used. Ij is a potential for obtaining an output, which is obtained from an input value to the unit.
X _j = f (I _j ) (5)
f (x) = 1 / (1 + e ^-x ) (6)
In the input layer, as shown in FIG. 18, data X (0) _j (j = 1 to N) at each sampling time (each data capturing position) is sequentially set in units of the input layer. The output value from each unit of the intermediate layer is determined using this value. For example, the j-th output value X (1) _j (j = 1 to M) of the intermediate layer is obtained from the following equation (7). N is the number of input values, and M is the number of hidden layers. Here, W (1) _ij is a weight coefficient between the i-th unit in the output layer and the j-th unit in the intermediate layer. Therefore, if the number of units in the input layer is N and the number of units in the intermediate layer is M, an M × N matrix is obtained. This matrix is called a weight matrix.
_N
X (1) _j = f (ΣW (1) _ij · X (0) _i ) (7)
i ^{= 1}
Similarly, the j-th output value X (2) ^j (j = 1 to L) of the output layer is obtained from Expression (8). L is the number of output layers. W (2) ^ji is a weighting coefficient between the i-th unit in the hidden layer and the j-th unit in the output layer. If the number of units in the hidden layer is M and the number of units in the output layer is L, L × M Is a matrix.
_N
X (2) _j = f (ΣW (2) _ji · X (0) _i ) (8)
^{i = 1}
In this case, L data is obtained by inputting N data. Generally, N, M, and L are arbitrary, but usually satisfy the relationship of N <M in many cases. The L pieces of data finally obtained as output values are data for a time of L × T _sam , where T _sam is the data sampling time. Therefore, when data correction is performed in real time, it is necessary to always input N data and obtain L output values within the time of L × T _sam . However, if the entire screen is read and all the image data is stored once, such restriction can be released. As the reading resolution increases, the number of data increases dramatically. Therefore, in any case, the calculation time needs to be high. For this reason, it is desirable that the operation unit is constituted by hardware. For example, when a digital signal processor of a pipeline processing type is used, the operation speed is higher than that of a conventional microprocessor.

As described in the overall estimating unit 55d, the weighting factor used in the above-described neural network is learned by using the actually correct data (teaching value) so that the distorted data becomes the same value as the teaching value as the output value. Go and make it the best. Many learning methods have been proposed, such as a back propagation method. The back propagation method is a learning method that minimizes an error relationship E defined as Expression 9 using an error between an output value and a teaching value in each unit of the output layer. Here, X (t) _i is a teaching value. However, any method may be used as long as the same value as the teaching value is finally obtained.
_L
E = (1/2) Σ (X (2) _i −X (t) _i ) ² (9)
^{i = 1}
In the case of a document reading apparatus as in the present embodiment, the cause of image data distortion is that image data is often modulated at a specific frequency. There are mainly three forms of this distortion. That is, the synthesis of the harmonic vibration, the AM modulation, and the FM modulation are expressed as Expression (10), Expression (11), and Expression (12), respectively. Expression (13) is a teaching value.
D ′ = φ ₀ F (t−τ ₀ ) + φ ₁ sin (ω ₁ t−τ ₁ ) (10)
D ′ = φ ₀ F (t−τ ₀ ) · φ ₁ sin (ω ₁ t−τ ₁ ) (11)
D ′ = φ ₀ F (t−τ ₀ · φ ₁ sin (ω ₁ t−τ ₁ )) (12)
D ′ = φ ₀ F (t−τ ₀ ) (13)
Here, ω ₁ is a disturbance frequency. Therefore, in the case of the learning of the weight Matorisuku the phase tau ₀ of the image data using the teachings value of formula (12), the phase tau ₁ of disturbance, maximum image density phi ₀ of data, the gain of the disturbance phi _1, disturbance Is the frequency ω ₁ .

First, an example of FM modulation in the case where φ ₀ and ω ₁ are fixed values will be described for easy understanding. FIG. 19 collectively shows the learning parameters τ ₀ , τ ₁ , and φ ₁ at that time. Using this value, learning is performed so that the output value is sufficiently close to the teaching value. Here, the values of the weight matrix are determined so that the errors are simultaneously minimized for each of the 32 cases.
The neural network does not necessarily need to learn in all cases. For example, if learning is performed when the phase of the disturbance is 30 ° and 60 °, a disturbance of 45 ° can be corrected to some extent. However, it is desirable to perform learning for as many cases as possible, because it is closer to ideal data correction.

FIG. 20 is a configuration diagram related to another embodiment of the present invention. In this figure, parts corresponding to those of the apparatus shown in FIG. 30 are denoted by the same reference numerals, and description of such parts is omitted. In this embodiment, a linear encoder 22 such as a magnetic scale is arranged along the movement path of the read head. The reading head 1 is provided with a sensor 21, reads an index of the linear encoder 22, and supplies a pulse train signal 24 carrying position information of the document 2 in the sub-scanning direction to the interface control unit 12. The interface control unit 12 supplies a pulse train signal to an internal counter reset at, for example, the home position or the opposite position of the read head 1, and determines the read position in the sub-scanning (Y) direction based on the count value. The scanning speed is determined by the increment. By monitoring the scanning speed, a change in the scanning speed is detected. The reading speed fluctuation information 25 and the position error information 26 that is the deviation of the reading position thus obtained are supplied to the image correcting unit 19 together with the image data, and are used for correcting the image data. Data and image data used for the correction are stored in the storage unit 20.

The speed fluctuation and position error of the read head obtained by the configuration using the sensor 21 and the linear encoder 22 are adopted as input data of the neural network. The learning uses the relationship between the speed fluctuation and the position error and the deviation of the image data obtained in advance through experiments and the like. In the case of this embodiment, it is basically necessary to use a large amount of previous data as in the above-described embodiment, since correction can be basically performed only by the image data itself and the speed fluctuation and position error of the time when it was obtained. Absent. If the speed fluctuation and position error of the direct read head cannot be obtained, the number of rotations of the motor is obtained by using a rotary encoder (not shown) or a rotation monitoring device and input, or an observer based on a modern control theory is installed and an appropriate value is set. Even if an input value is obtained, the effect of the present invention can be obtained. Therefore, the calculation for obtaining the output value using the weight matrix is simplified, and the calculation time can be reduced.
In an image correction method using a neural network, a predetermined output is obtained using multiple inputs. For this reason, when the input values include the use environment such as temperature and humidity in addition to the embodiment, the data corresponding to the state of the mechanism can be corrected according to the use environment.

When an image correction method using a neural network is used, an embodiment described below can be configured.
(1) Image area separation
One document read by the image reading device may include both a graphic portion and a character portion. In such a case, the image data can be corrected using different neural networks for the graphic part and the character part. FIG. 21 is a flowchart for explaining an algorithm in such a case, in which image data read by the image reading device is temporarily stored in a storage device (steps S61 and S62). For one frame of the image data, for example, an image area separation process is performed using a pattern recognition algorithm (not shown) to divide the image data into a character data area in which characters are extracted and a graphic area in which the characters are not extracted (step S63). The image data (graphic data) existing in the graphic area in one frame is provided to the graphic part neural network to correct the image data (step S64). The image data (character data) existing in the character data area in one frame is supplied to the character part neural network to correct the image data (step S65). The two types of corrected image data are combined and returned to the original of one frame (step S66). The combined image data is stored in the storage device (step S67).
As described above, when correcting one frame of image data, by appropriately selecting and using a neural network corresponding to the type or property of the image data in one frame, the accuracy of the data supplement is further improved. We can expect that.

(2) Vibration area separation
In some cases, distortion of image data in the image reading device occurs at a specific location in the image reading area. For example, when the guide rail of the read head is dirty or scratched. As shown in FIG. 7, the jitter as the time axis fluctuation can also be obtained by comparing the unit wavelength of the density curve obtained by using the test pattern with the reference wavelength. In the image reading apparatus shown in FIG. Is obtained directly by the sensor 21, so that the detection of the speed fluctuation is easier. The image data can be corrected only for the image data in the area where the speed fluctuation has occurred.

FIG. 22 is a flowchart for explaining an algorithm in such a case, in which a pre-scan is performed in advance to detect a vibration occurrence location, and image data corresponding to the vibration occurrence location among image data read later. And correct it.
The carriage that carries the read head is started from the home position to the opposite position (step S71). At the same time as the start, the speed fluctuation ΔV in the sub-scanning direction is measured (step S72). When the speed fluctuation ΔV exceeds the predetermined reference value Vct, this position is determined, and a range in which vibration occurs in the scanning direction is stored (step S73). There may be a plurality of places where vibration occurs. In this case, image data correction is performed for each occurrence location. Such an error occurrence area can be determined, and a prescan for storing the location can be automatically executed immediately after the power of the apparatus is turned on.

(4) The document image is read and stored in the image memory (step S74). Among the read image data, the image data in the error occurrence area is separated and moved to the correction data storage area. Note that while reading the image, the read image data can be separately stored in the image data of the error occurrence area and the image data of the other areas (step S75).

(4) The image data of the error occurrence area stored in the correction data storage area is corrected using a neural network (step S76). The image data corrected in the error occurrence area and the image data in the other areas are combined to form one frame of image data (step S77).

In this manner, the processing time can be reduced because the amount of image data to be corrected is reduced as compared with the case where image data correction is performed on the entire area of an image of one frame, and the image data is stored when processing is performed. There is an advantage that the memory capacity can be small.

(3) Using multiple neural networks Character documents, graphic documents, color documents, monochrome documents, documents described on document paper, front documents, documents containing frames, halftone documents, etc. By individually selecting the neural networks correspondingly, it can be expected that the accuracy of the correction of the image data is further improved.

{Circle around (1)} First, an original is read, an optimal image correction neural network is selected by the original determination neural network, and the read image data is corrected using the selected neural network. In a neural network, as the number of neurons increases, the time required for output increases, and the capacity of a memory for storing a weight matrix also increases. Therefore, it is desirable to reduce the number of inputs to the neural network as much as possible by using image data read at a resolution lower than the normal resolution or extracting features from the image data of the document as the image data used for document determination. There are various methods for extracting features for document determination, in addition to using a neural network. For example, calculating the density gradient of some edges on the document, using this method for determination, comparing the distribution of the frequency spectrum of the read image data with the distribution of the frequency spectrum of the document pattern known in advance, A method of estimating the type of the document from the degree of approximation is used. In order to easily perform the document determination, an operator is provided with a document document selection button for designating the type of the document on the operation panel of the apparatus, and the corresponding neural network is selected in accordance with the designation by the button.

FIG. 23 shows an algorithm when a plurality of neural networks are used. First, in order to determine the type of the document, the image data is read by pre-scanning and stored in a memory. As described above, the read pixels at this time can be made coarse, and only a predetermined area of the document may be sampled (step S81). Using this image data, feature parameters representing the features of the document such as density gradient, edge extraction, image area ratio, average density, solid black discrimination, color component extraction, line segment length, curve length discrimination, etc. are extracted ( Step S82). The characteristic parameters are given to the original determination neural network (NN) to determine the type of the original. An appropriate image correction neural network is selected according to the determination result. In other words, the corresponding one is read from the storage device from the file of the weight matrix for each document type of the neural network, and is set in the arithmetic unit of the neural network (step S83). The scanning is started to read the original, and the image data is stored in the memory (step S84). The image data is corrected using the image correction neural network (step S85).

(4) Processing direction of image data correction
When the image data is corrected by the neural network, the manner in which the image data is given to the neural network may be any of the main scanning direction and the sub-scanning direction of the read image data, and is appropriately selected.

FIG. 24 shows a case where the correction of the image data is repeated in the main scanning direction by the neural network, and the image data of the first example composed of i pixels in the sub-scanning direction is corrected corresponding to the number of input terminals i of the neural network. After the processing, the correction processing is repeated for the second example, the third column,..., The m-th column, and the correction processing is similarly repeated from the first column of the next row direction data block.

FIG. 25 shows a case where the correction of the image data is repeated in the sub-scanning direction by the neural network, and the image data of the first row composed of 1 pixel in the main scanning direction is corrected corresponding to the number of input terminals 1 of the neural network. After the processing, the correction processing is repeated for the second row, third row,... N row, and the correction processing is similarly repeated from the first row of the next column direction data block.

(5) Inspection tool
The present invention can be used as an image quality inspection tool. When image data obtained by reading a specific test chart is given to a neural network that has been subjected to backpropagation learning so that a predetermined output can be obtained from the test chart, an output depending on the current error state of the apparatus is generated. I do. Therefore, for example, it is possible to read a test chart having the same pitch and determine “adjustment failure (adjustment required)”, “adjusted (adjustment unnecessary)” and the like. In the case of adjustment failure, the cause can be further estimated using a self-diagnosis neural network.

When the image reading device has a self-healing function, an optimum self-healing operation is commanded and executed based on the result of the self-diagnosis of the state of the apparatus, so that the state of the apparatus can always be maintained in a good state. . For example, when it is diagnosed that the fluctuation of the rotation speed of the motor used is large when the speed fluctuation is large, the device itself performs an operation of appropriately changing the constant in the servo controller of the motor to reduce the rotation speed fluctuation. You can do it. It is also possible to apply fuzzy control to the determination of the control amount for the motor. An expert system can be applied to determine the self-healing operation. The learned knowledge can be loaded with the one developed offline before shipping the device.

FIG. 26 shows a configuration example of such a system. In the figure, image data, scanning speed, image density, image reading position, and the like are obtained by the image reading device (step S111). A state quantity representing a state of the apparatus such as generation of vibration is extracted (step S112). The extracted state quantity is given to, for example, a state discrimination neural network, and it is determined whether or not a problem has occurred, its degree, and the like (step S113). If a defect that affects the read image quality occurs, the cause is estimated based on the content of the defect and various state quantities. Determine the coping method according to the estimated cause. This determination can be made using the expert system described above. If the self-repair is not sufficient, a process such as issuing a warning to the operator is performed (step S114). In order to solve the problem, a control parameter of each unit of the image reading apparatus that needs to be operated is selected. For this selection, a neural network can be used (step S115). Change the setting of the selected control parameter. For example, when the read head is vibrating, the loop constant of the servo motor is set so that the rotation fluctuation is further reduced. When there is a change in the light amount, adjustment of the supply voltage to the light source and adjustment of the gap between the platen glass and the reading head are performed. If dirt on the guide rail of the read head is detected, the operation interval of the rail cleaning mechanism is shortened. In addition, when there are a plurality of control targets for one fault, or when there are a plurality of faults, it is possible to perform approximate control together with fuzzy calculation in order to determine a control amount to be operated. is there. In such a case, there is an advantage that the control becomes simpler and a control error is less generated as compared with a complicated system in which the proportional control is complicated (step S116).
In this way, the arithmetic mechanism for the neural network incorporated for correcting the image data can be used for multiple purposes.

(6) Use of spectrum of image data When the spectrum in the sub-scanning direction of the image data is calculated by the test chart in which the lines of equal pitch are printed, the peaks appear include those caused by the vibration of the device in the sub-scanning direction. . Also, the value of the original peak generated corresponding to the above equal pitch may be reduced by disturbance. The peak value is related to the amplitude of the vibration. Therefore, a method is conceivable in which learning is performed by using this peak value as an input, and the influence of the disturbance vibration on the image data is removed by a neural network. However, depending on the case, there is a possibility that even the frequency components inherent in the original document may be deleted. For this reason, it is necessary to estimate the peak due to the disturbance and to accurately examine the influence on the frequency components originally possessed by the original. Therefore, assuming in advance all disturbances and frequency components of the original, inputting the image data of the original including the disturbance components (gain and phase) and outputting the disturbance components appearing in the spectrum as much as possible. Train with many patterns. As a result, a disturbance component included in any document data can be extracted.

(5) Another neural network is trained to remove a disturbance component from image data containing a specific disturbance component. Then, parameter settings of the neural network capable of removing a specific disturbance component are stored for various disturbance components. Using these two types of neural networks, it is possible to remove disturbance components from an arbitrary document.

However, there may be cases where it is difficult to apply it to any disturbance component. In this case, a practical effect is large even if the method is limited to a method in which a frequency component (for example, due to a motor) that can be assumed is fixed in advance and disturbance is removed at that frequency. This is because the cause of vibration in the device is usually limited to a certain range due to the mechanical structure of the device. Conventionally, disturbances have been removed in an open loop as described above, so that accurate disturbances cannot be removed. The benefit of removing this disturbance is significant. The flow of the above-described processing is shown in FIGS.

When the disturbance component is constantly generated, prescanning is performed as shown in FIG. 27 to read the original (step S121), and the image data of the original is supplied to the disturbance component detecting neural network to remove the disturbance component. It is extracted (step S122). The image data used for detecting the disturbance component may be only a part of the document. The parameters are set in the disturbance component removing neural network so as to remove the disturbance component. The whole original is read (step S123), and the image data is supplied to a disturbance component removing neural network to perform data correction for removing disturbance components from the image data (step S124). When the data is corrected for the entire area of the document, the data of the read document can be used repeatedly as shown in steps S131 to S133 in FIG. 28, so the document reading operation may be performed only once.

Although the embodiment has been described in connection with the case where noise is mixed in the image data when the reading head scans, for example, the document may be read by flash exposure using a CCD having a two-dimensional array of reading elements. The present invention can be applied to a reading device having a simple configuration. In such an apparatus, mechanical vibration is usually slight, but so-called pincushion distortion, barrel distortion, and local projection image distortion occur in an optical system for projecting an original image onto a CCD. As described above, the data correction using the neural network is also effective for the error correction of the image data due to the cause.

・ Density correction method based on CCD charge estimation
As described above, the conventional image reading apparatus often uses a data acquisition circuit that samples the density information from the CCD at regular time intervals. In this case, the moving speed of the reading head needs to be constant in order to spatially disperse the density information of the document at a constant distance. However, if a higher resolution is required in the future, it is expected that it will be difficult to suppress the fluctuation of the moving speed of the read head within a predetermined specification. A method for solving this problem will be described below.

As shown in FIG. 20, in an image reading apparatus provided with encoders (21, 22), data can be sampled in synchronization with the position of the reading head 1. In this case, if there is a change in the moving speed of the reading head 1, the data acquisition (image projection) time in the CCD 4 also changes. The CCD 4 basically associates the charge amount charged in proportion to the light intensity and the irradiation time with the image density value. Therefore, even if the original is printed at the same density, if the scanning speed of the read head 1 fluctuates, the amount of charge charged due to the fluctuation of the data acquisition time in the CCD 4 is not exactly proportional to the density of the original. An error occurs.

In such a case, the read image density data can be corrected to a correct density value using the moving speed fluctuation amount of the reading head 1 obtained by the sensor 21 and the encoder 22. That is,
In a CCD, a voltage obtained as an amount proportional to the image density is used as a density value. The output voltage y of the CCD is generally represented by the following equation.
y = al _x t + b (13)
Where a is the sensitivity, b is dark output voltage, l _x illuminance, t is the charge storage time of at the CCD. l _x corresponds to the amount of reflection of light impinging on the document by the lamp. Therefore, y is negative data of the density information of the document. Now, when a certain density data and y ₁ accumulation time t ₁ at this time is represented by the following formula (14).
t ₁ = x ₀ / v ₁ (14)
Here, x ₀ is a moving distance, v ₁ is the moving speed.
Therefore, the correct density data y ₀ when there is no speed fluctuation is
y ₀ = al _x t ₀ + b (15)
t ₀ is a correct sampling time when there is no speed fluctuation.
From Expressions (14) and (15), correct density data y ₀ can be described by y ₁ and t ₁ as in Expression (16).
y ₀ = t ₀ / t ₁ × (y ₁ −b) −b (16)
t1 may be obtained by using equation (14) or by measuring the time interval between pulses from the encoder. That is, if the speed at which the image data was obtained is known, it can be corrected to the correct density data using Expression (16). Equation (13) has a so-called linear relationship between the density data and the accumulation time. However, even in the case of non-linearity or when expressed by some function, the correction by the speed and the accumulation time described above is possible here. It is.

The determination of whether or not to perform image density correction by estimating the charge amount of the CCD is performed by sampling image data in synchronization with the reading scanning position, and when the data acquisition time in the CCD changes due to the speed fluctuation of the reading head, the image data is determined. Since a spectrum appears at a random frequency in the obtained image error spectrum distribution, a circuit for performing the above-described image correction is started in such a case. Also, even when such density correction has already been performed by estimating the charge amount of the CCD, the sensitivity a and the dark output voltage b in the equation (13) change with temperature and aging. If the adjustment is performed by the image correction algorithm when performing the correction, the optimum image correction according to the characteristics of the CCD can always be performed.

(4) Manual adjustment unit (flag K = 5)
In the description of the motor control unit, those related to the manual adjustment unit have been described. However, the data may not be improved even if the motor control and the image correction are performed. In such a case, if the cause can be identified to some extent in such a case, the location of the problem and a countermeasure method are instructed, and the assembly adjustment is supported.

From the information obtained by the image data error identification circuit, the following relationship is known.
-Correlation with the natural frequency of each mechanism obtained from the spectrum peak.
-The inherent vibration of the reducer obtained from the spectral peak.
-Optical system errors (dirt, defocus) obtained from the spectral peaks.
-Lubrication failure without sliding obtained from the reading position error distribution.
-Insufficient light intensity of the lamp due to a decrease in density.
-Dirt on the platen that can be seen from the density data distribution.
In addition, by incorporating empirically obtained correlations into an inference method such as an ES (expert system), an appropriate countermeasure can be selected from information obtained by the image data error identification circuit.

(5) Other information indicating section (flag K = 6)
In the information extraction from the image data or the like in the image data error feature extraction unit 54, information on a purpose other than the adjustment is also obtained. Here, the information is output. An example of information handled here is shown below.
-Frequency histogram of image density.
・ Reproducibility of reading position error.
-Peak value and band value of each spectrum peak.

(Information display section)
The information obtained by the image data error feature extraction unit 54 and the adjustment method determination unit 56 described above is displayed on a display set in an appropriate part of the apparatus. Alternatively, the information may be displayed on an external output device via an interface.

In the case of inputting a color image, the image data color-separated into RGB, YMC, etc. basically corresponds to the intensity of the reflected light of each wavelength component, so that the apparatus according to the present invention can be adjusted. . Also, using the reading position error distribution, it is possible to confirm and adjust the performance of items unique to a color image, such as position reproducibility and color tone of each color.

FIG. 29 shows a configuration example as a more general image input device that evaluates errors and image quality in image data and adjusts an image input unit and an image output unit. In the figure, an image input unit 291 composed of an electronic camera, a scanner, or the like captures an image or scans a document to read an image, and sends this to the image processing unit 292 as image data. The image processing unit 292 arranges the image data into a predetermined format and sends the data to the image output unit 293. The image data is also sent to the image evaluation unit 294. The image data to the image evaluation unit 293 does not necessarily need to be image data of one entire frame, but may be only data of a certain evaluation image area. The image evaluation unit 293 has the same configuration and functions as the above-described image data error feature extraction unit 54, error factor estimation unit 55, and adjustment method determination unit 56. Is evaluated. If an error factor or image quality is related to the image input stage, the image input unit adjustment unit 295 adjusts the image input unit 291. For example, when it is determined that the level of the image data is low, the gain of the preamplifier of the CCD for reading the image is adjusted. In addition, based on the evaluation, the image output unit adjustment unit 296 controls the image output unit 293 to correct the read image data, and corrects the output image data. If the output image is fed back, the adjustment converges to a constant state.

画像 The image input device having such a configuration performs self-diagnosis based on the supplied image data, grasps its own characteristics, and sufficiently adjusts itself. This image input device can be used for performance evaluation and adjustment assistance of another image processing device. If the characteristics of the image error and the state of the device for an apparatus to be evaluated, for example, an electronic camera, a facsimile, a digital copying machine, etc., are set in advance in the error factor estimating portion of the image evaluation section 294, the adjustment of the electronic camera, etc. Can be performed. Such a function is particularly effective in a digital copying machine that reads an image, processes the read image, and outputs the processed image to a printer.

According to an embodiment of the present invention, in an image reading apparatus that captures image information while the reading head moves, errors in image data such as fluctuations in the moving speed of the reading head, abnormal vibration of the mechanism of the device, and displacement of the optical system. Even if a problem occurs, the error information is extracted by the image data error feature extraction unit, the error factor is determined by the error factor estimation unit, and the corresponding part is adjusted by the adjustment method determination circuit to eliminate the error factor. Then, the read image data is corrected by selecting a correction method according to the characteristics of the error. As a result, image data close to the true one is obtained. For this reason, specifications required for the mechanical performance of the apparatus can be easily realized, and the cost of the apparatus can be reduced. Furthermore, the reading resolution performance of the scanner itself is improved by the error correction.

FIG. 1 is a system block diagram of an image reading apparatus according to an embodiment of the present invention. 9 is a flowchart illustrating a data processing procedure in the image data error feature extraction circuit. FIG. 4 is an explanatory diagram illustrating an example of a test chart in which lines are printed at equal pitches. 5 is a graph showing an example of a density distribution of extracted image data. 7 is a graph showing an example of a density distribution of data obtained by normalizing extracted image data. 5 is a graph showing an example of a density distribution of interpolated image data. FIG. 4 is a view for explaining the definition of a reading position error (jitter). 5 is a graph illustrating an example of a distribution of a reading position error in a sub-scanning direction. 7 is a graph showing an example of a spectral distribution of a reading position error. FIG. 3 is a block diagram illustrating a configuration example of an error factor estimating unit. FIG. 3 is a block diagram illustrating a configuration example of an adjustment method determination circuit. 5 is a graph illustrating an example of a distribution of a reading position error in a sub-scanning direction. 7 is a graph showing an example of a spectral distribution of a reading position error. 5 is a graph illustrating an example of a distribution of a reading position error in a sub-scanning direction. 5 is a graph illustrating an example of a density distribution of image data in a Y direction. 5 is a graph illustrating an example of a density distribution of image data including noise. FIG. 3 is an explanatory diagram showing a structure of a neural network used for correcting image data. FIG. 4 is an explanatory diagram showing the correspondence between input values and units in an input layer. The figure which shows the example of the parameter at the time of learning of a neural network. FIG. 2 is a block diagram illustrating an example of an image reading apparatus capable of directly monitoring the speed and position of a reading head. 9 is a flowchart illustrating an algorithm when correcting image data by performing image area separation. 9 is a flowchart illustrating an algorithm when image data is corrected by performing vibration area separation. 9 is a flowchart illustrating an algorithm when a plurality of neural networks are used. FIG. 4 is an explanatory diagram illustrating an example of a case where image correction is performed in a main scanning direction by a neural network. FIG. 9 is an explanatory diagram showing an example of a case where image correction is performed in the sub-scanning direction by a neural network. FIG. 2 is a block diagram showing a configuration example when the present invention is used as an inspection tool. 9 is a flowchart illustrating an example of a case where a spectrum of image data is used as disturbance detection. 9 is a flowchart illustrating an example of a case where a spectrum of image data is used as disturbance detection. FIG. 2 is a block diagram related to the present invention. FIG. 9 is a block diagram illustrating a configuration example of a conventional image reading apparatus.

Explanation of reference numerals

1 read head 2 original 3 platen (glass)
4 CCD
5 Optical system 6 Image reading direction 7, 17 A / D converter 8 Motor 9 Pulley 10 Belt 11, 20 Storage device 12, 18 Interface controller 13 Motor controller 14 Motor drive unit 15 Lamp 16 Image reading device 19 Image correction unit 21 Sensor 22 Linear encoder 23 Image data 24 Pulse train 25 Speed fluctuation value 26 Position error 27 Test chart 53 Mode selection unit 54 Image data error feature extraction unit 55 Error factor estimation unit 56 Adjustment method determination unit 57 Image correction unit

Claims (1)

By reading the test chart, errors on the time axis and the frequency axis are obtained, a first error as an error on the time axis and a first error factor as a factor thereof, and an error on the frequency axis are obtained. Is stored in advance in the first database, and a second error as
Thereafter, an image of the target document is read, and the first error and the first error factor in the first database are determined based on data on the time axis and the frequency axis obtained from the read image data. And the corresponding first and second error factors are picked up from the corresponding relationship between the second error factor and the second error factor,
And executing a correction algorithm for canceling an error occurring in the image data based on the first and second error factors picked up.
An image correction method, comprising:

Images