Image forming apparatus, method, and program

Abstract

To improve accuracy of image recognition.SOLUTION: An image formation device being one embodiment of the invention comprises: a binary image generation unit which generates a binary image having area gradation properties, from an image scanned with a scanner; and an image recognition unit which outputs classification according to a neural network model learned in advance, to the generated image.SELECTED DRAWING: Figure 11

Description

Technical Field

[0001] The present invention relates to an image forming apparatus, method, and program.

Background Art

[0002] Conventionally, when reading a document with a scanner, the document may be read with the top and bottom reversed or in a landscape orientation (for example, what should be read at "north" in FIG. 1 may be read at "south" (that is, with the top and bottom of the document reversed), or at "east" or "west" (that is, with the document in a landscape orientation)).

[0003] For example, there is a technique (hereinafter also referred to as "top-bottom discrimination") for automatically discriminating the top and bottom of a document read in such a state and correcting it to a normal orientation (Patent Document 1). As a conventional method, a method of performing top-bottom discrimination using OCR (Optical Character Recognition) is known.

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, the method using OCR has a problem that it is strong for documents containing characters, but it is difficult to handle documents such as photographs that contain almost no characters.

[0005] Therefore, an object of the present invention is to improve the accuracy of image recognition.

Means for Solving the Problems

[0006] An image forming apparatus according to an embodiment of the present invention includes a binary image generation unit that generates a binary image having area gradation from an image read by a scanner, (1 bit / pixel) and an image recognition unit that outputs a class classification according to a neural network model learned in advance for the above. A grayscale conversion unit that converts the aforementioned binary image (1 bit / pixel) to a binary image (8 bits / pixel), For the above Binary image (8 bits / pixel) The image forming apparatus is provided with an image recognition unit that outputs a class classification according to a neural network model learned in advance.

[0007] According to the present invention, the accuracy of image recognition can be improved. [Brief explanation of the drawing]

[0008] [Figure 1] This is an example of the scanning orientation for the original document. [Figure 2] This is an input image (category) suitable for each method. [Figure 3] The input image (binary or multi-level image) is suitable for each method. [Figure 4] This is an example of an input image related to one embodiment of the present invention. [Figure 5] This is an example of a binary image that does not maintain grayscale according to one embodiment of the present invention. [Figure 6] This is an example of a binary image that maintains grayscale according to one embodiment of the present invention. [Figure 7] This is an example of a scaled-down image (reduced image) that maintains grayscale according to one embodiment of the present invention. [Figure 8] This is a diagram illustrating the processing flow related to one embodiment of a particular model. [Figure 9] This is a diagram illustrating the processing flow related to one embodiment of a particular model. [Figure 10] This is a diagram illustrating the processing flow related to one embodiment of a particular model. [Figure 11] This is a functional block diagram of an image forming apparatus according to one embodiment of the present invention. [Figure 12] This is a functional block diagram of an image processing unit according to one embodiment of the present invention. [Figure 13] This is a flowchart of the image processing and orientation recognition processes related to one embodiment of the present invention. [Figure 14] This is a diagram illustrating the image data binarization process according to one embodiment of the present invention. [Figure 15] This diagram illustrates the image data scaling process according to one embodiment of the present invention. [Figure 16] This is a hardware configuration diagram of an image forming apparatus according to an embodiment of the present invention.

Embodiments for Carrying Out the Invention

[0009] Hereinafter, embodiments of the present disclosure will be described based on the drawings.

[0010] In this specification, the case of top-bottom discrimination (that is, automatically discriminating the top and bottom of a document and correcting it to a proper orientation) is described as an example, but the present invention can be applied not only to top-bottom discrimination but also to any discrimination and classification such as form discrimination and document type discrimination.

[0011] <Summary> A conventional method of performing top-bottom discrimination using OCR has a problem in that it is strong for documents containing characters but difficult to handle documents such as photos that contain almost no characters. Therefore, in one embodiment of the present invention, top-bottom discrimination is performed using DeepLearning technology.

[0012] A method of performing top-bottom discrimination using DeepLearning technology has an advantage for documents such as photos without characters because it does not use OCR. Fig. 2 shows these relationships. As shown in Fig. 2, an image suitable for top-bottom discrimination using OCR is a character image (an image containing characters), and an image suitable for top-bottom discrimination using DeepLearning is a character image and a natural image (an image without characters or an image containing almost no characters).

[0013] OCR is a method premised on using binary images, but in DeepLearning image recognition, it is common to use multi-value images. Fig. 3 shows these relationships. As shown in Fig. 3, an image suitable for top-bottom discrimination using OCR is a binary image, and an image suitable for top-bottom discrimination using DeepLearning is a multi-value image.

[0014] However, when using multi-valued images, there is a tendency for the memory consumption and the like to increase more than when using binary images, and depending on the conditions of the device equipped with the model learned by DeepLearning, it may be difficult to prepare multi-valued images.

[0015] As described above, multi-valued images are preferable when using DeepLearning, but due to device constraints, it is often difficult (that is, it is desirable to be able to process with binary images). However, in simple processes (binarization, resizing (shrinking and enlarging)), models based on CNN (Convolutional Neural Network) tend to lose image features during the calculation process, and the recognition rate tends to be low.

[0016] Here, the above-mentioned CNN generally has a configuration that repeats filters and thinning, and the filter outputs a multi-valued image even when a binary image is input. That is, in the case of DNN (Deep Neural Network) using a filter such as CNN, even if the input is a binary image, the subsequent image becomes multi-valued and the subsequent processing is performed. Therefore, it is ideal if the multi-valued image after filtering can be made like one that has input a multi-valued image from the beginning even when the input image is binary.

[0017] Also, even when processing such as resizing (shrinking and enlarging) is performed before directly inputting a binary image into the NN, it is ideal if the multi-valued image after filtering can be made like one that has input a multi-valued image from the beginning.

[0018] Therefore, in the present invention, by performing processing that leaves gradation as much as possible like a multi-valued image, a recognition rate equivalent to that of a multi-valued image is realized even when using a binary image.

[0019] Figures 4-7 show examples of images with maintained gradation, along with images that have undergone simple processing without maintaining gradation. Figure 4 is an example of an input image (multi-level image) according to one embodiment of the present invention, Figure 5 is an example of a binary image without maintained gradation according to one embodiment of the present invention, Figure 6 is an example of a binary image with maintained gradation according to one embodiment of the present invention, and Figure 7 is an example of a scaled image (reduced image) with maintained gradation according to one embodiment of the present invention.

[0020] Figure 8 is a diagram illustrating the processing flow related to one embodiment of the present invention.

[0021] <Learner side> As a preprocessing step, (1) is performed to binarize the image (to create a binary image that takes area gradation into account (e.g., error diffusion method)). Figure 8 also shows the training process, which is an example of applying the same preprocessing as during inference. However, in the case of the present invention, any pre-trained model can be used if the input is a grayscale image. Next, in (2), set the bit depth to 8 bits / pixel (if it is 1 bit / pixel). Note that (1) and (3) may be performed simultaneously. Next, in (3), the image is scaled down (to create a scaled-down image with tonality (e.g., area averaging method, Gaussian filter + bicubic method, etc.)). Next, in (4), the image is trained (as a multi-level image).

[0022] <Inference side> As a preprocessing step, (1) is performed to binarize the image (to create a binary image that takes area gradation into account (e.g., error diffusion method)). Figure 8 also shows the training process, which is an example of applying the same preprocessing as during inference. However, in the case of the present invention, any pre-trained model can be used if the input is a grayscale image. Next, in (2), set the bit depth to 8 bits / pixel (if it is 1 bit / pixel). Note that (1) and (3) may be performed simultaneously. Next, in (3), the image is scaled down (to create a scaled-down image with tonality (e.g., area averaging method, Gaussian filter + bicubic method, etc.)). Next, in (4), we perform inference (as a multi-level image).

[0023] Figures 9 and 10 are diagrams illustrating the processing flow of one embodiment of the present invention.

[0024] First, let me explain each image. In this specification, a "binary image (8 bits / pixel)" is an image in which the value of each pixel is two colors (i.e., two colors, black and white) and the color has multiple gradations.

[0025] Branch (1) in Figure 9 refers to the first branch counting from the start of Figure 10, "Is the input image data a binary image?". Branch (2) in Figure 9 refers to the second branch counting from the start of Figure 10, "Is the bit depth of the input image 1 bit / pixel?". Branch (3) in Figure 9 refers to the third branch counting from the start of Figure 10, "Is the size of the input image data a pre-set size?".

[0026] As shown in Figure 9, No. 1, if the input image at branch (1) is a multi-level image (8 bits / pixel) (i.e., a color image or a grayscale image), the binarization process outputs a binary image (1 bit / pixel) and proceeds to branch (2). At branch (2), since it is a binary image (1 bit / pixel), it is converted to a binary image (8 bits / pixel) and proceeds to branch (3).

[0027] As shown in Figure 9, No. 2, if the input image from reference (1) is a binary image (8 bits / pixel), proceed directly to branch (2). At branch (2), since it is a binary image (8 bits / pixel), proceed directly to branch (3).

[0028] As shown in Figure 9, No. 3, if the input image at branch (1) is a binary image (1 bit / pixel), the process proceeds directly to branch (2). At branch (2), since it is a binary image (1 bit / pixel), it is converted to a binary image (8 bits / pixel) and the process proceeds to branch (3).

[0029] Thus, while it is preferable to prepare multi-level images when employing Deep Learning technology for orientation discrimination, it is possible to achieve orientation discrimination accuracy equivalent to that of multi-level images even in environments where only binary images are available due to device limitations. A Deep Neural Network (DNN), which uses filters like those used in Convolutional Neural Networks (CNNs) as its constituent elements, has the characteristic that even if the input is a binary image, the subsequent image becomes multi-level for later processing. By utilizing this characteristic, pre-processing such as binarization and scaling is performed on the DNN input in a way that preserves (as much as possible) the gradation, similar to a multi-level image. As a result, the input image to the DNN becomes as if a multi-level image had been input from the beginning, and recognition accuracy equivalent to that of a multi-level image can be expected.

[0030] <Function Block> Figure 11 is a functional block diagram of an image forming apparatus (digital color image forming apparatus) 100 according to one embodiment of the present invention. The digital color image forming apparatus 100 comprises a scanner 1, an image processing device 2, an HDD 3, a plotter 4, and an image file format conversion device 5. Furthermore, the digital color image forming apparatus 100 functions as the scanner 1, image processing device 2, HDD 3, plotter 4, and image file format conversion device 5 by executing a program.

[0031] Scanner 1 is a device that reads image data from a document. Scanner 1 sends the read image data to image processing device 2.

[0032] The image processing device 2 stores in the area detection unit 22 a character determination unit that determines whether a pixel or pixel block of interest in the image read by the scanner 1 is a character area or a non-character area (i.e., a picture), and in the color determination unit that determines whether it is chromatic or achromatic, and performs color reproduction processing suitable for the original in the color processing / UCR unit 24 according to the result.

[0033] The plotter 4 is a transfer printing unit. The plotter 4 transfers the image data output by the image processing device 2.

[0034] The image processing device 2 consists of a gamma correction unit 21, a region detection unit 22, a data interface unit 23, a color processing / UCR unit 24, and a printer correction unit 25, and performs the process of obtaining a copy image.

[0035] The gamma correction unit 21 is a means of applying a one-dimensional conversion to each signal in order to equalize the gradation balance of each color in the data read by the scanner 1 (8 bits each for r, g, and b after A / D conversion). For the purpose of this explanation, the converted signal is a density linear signal (RGB signal: the signal value representing white is set to 0). The output of the gamma correction unit 21 is sent to the area detection unit 22 and directly to the data interface unit 23.

[0036] The data interface unit 23 is an HDD management interface for temporarily saving the judgment results from the region detection unit 22 and the image data processed by the gamma correction unit 21 to the HDD 3. Meanwhile, the image data processed by the gamma correction unit 21 and the judgment results from the region detection unit 22 are sent from the data interface unit 23 to the color processing / UCR unit 24.

[0037] The color processing / UCR unit 24 is a means for selecting color processing or UCR processing based on the determination result for each pixel or pixel block.

[0038] The printer correction unit 25 receives the c, m, y, and Bk image signals from the color processing / UCR unit 24, performs gamma correction and dithering processing considering the printer characteristics, and then sends the image to the plotter 4.

[0039] The image file format conversion device 5 receives the image data processed by the gamma correction unit 21, which has been temporarily stored on the HDD 3, and performs orientation recognition. Using the result obtained from orientation recognition, the image file format conversion device 5 converts the image to a file format (format) such as PDF or the Office Open XML Document format used in Microsoft Word.

[0040] The image file format conversion device 5 consists of an image processing unit 51, an orientation recognition unit 52, and a file format conversion unit 53, and performs processing for orientation recognition (more precisely, it performs orientation recognition and then converts the file format using the orientation recognition result).

[0041] The image processing unit 51 applies binarization and scaling to the image data processed by the gamma correction unit 21. The image data converted by the image processing unit 51 is output to the orientation recognition unit 52.

[0042] The orientation recognition unit (also called the image recognition unit) 52 performs orientation recognition by inputting the image output from the image processing unit 51 into a pre-trained recognition model. The inference (recognition) result (or orientation recognition result) will be one of the following: north, east, west, or south. The orientation recognition result obtained by the orientation recognition unit 52 is output to the file format conversion unit 53.

[0043] Here, we will describe the recognition model. In one embodiment of the present invention, the recognition model is a neural network model configured in multiple stages with multiple filters. In another embodiment of the present invention, the recognition model is a neural network model trained using grayscale images as input.

[0044] Note that the example shown uses north, east, west, and south as the correct labels (i.e., the orientation of the image), but there are no restrictions; any index from 0 to 3 is acceptable as long as the relationship between the orientation and the label matches. Also, there is a correlation between the image and the correct label, and the relationship between the image and the correct label does not change depending on the subject.

[0045] The file format conversion unit 53 utilizes the orientation identification result output from the orientation identification unit 52 to convert the file into a file format (format) such as PDF or the Office Open XML Document format used in Microsoft Word.

[0046] Figure 12 is a functional block diagram of an image processing unit 51 according to one embodiment of the present invention.

[0047] The image processing unit 51 consists of a binary image generation unit 511, a grayscale conversion unit 512, and a variable-size image generation unit 513.

[0048] The image processing unit 51 performs image processing (conversion) such as binarization and scaling on the image data input from the HDD 3, and the converted image data is input to the orientation recognition unit 52.

[0049] The binary image generation unit 511 performs a binarization process on the image input from the HDD3 based on a binarization algorithm (such as the error diffusion method) that takes area gradation into consideration. If the image input from the HDD3 is a multi-level image (color image or grayscale image), the binary image generated by the binarization process is output and input to the grayscale conversion unit 512. If the image input from the HDD3 is a binary image, the binary image generation unit 511 does not perform any processing on the input image, and the image input from the HDD3 is input directly to the grayscale conversion unit 512.

[0050] The grayscale conversion unit 512 converts the binary image (1 bit / pixel) input from the binary image generation unit 511 into a binary image (8 bits / pixel) in a format suitable for subsequent processing. The binary image (8 bits / pixel) converted by the grayscale conversion unit 512 is input to the variable image generation unit 513. If the image input from the binary image generation unit 511 is already a binary image (8 bits / pixel), it is input to the variable image generation unit 513 without any special conversion processing.

[0051] The resizing image generation unit 513 performs resizing processing on the binary image input from the grayscale conversion unit 512 based on a resizing algorithm that takes area gradation into consideration (area averaging method, Gaussian filter + bicubic method, etc.) so that the image size becomes the input size required by the subsequent recognition model (for example, the recognition model for orientation recognition used in the orientation recognition unit 52). At this time, if the image size of the binary image input from the grayscale conversion unit 512 is larger than the input size required by the subsequent recognition model, the resizing process becomes a reduction resizing process, and if it is smaller than the input size required by the subsequent recognition model, it becomes an enlargement resizing process (therefore, the resizing ratio is uniquely determined by the image size of the input image). The resized image (resizing image) of the binary image input from the grayscale conversion unit 512, which has been resized to the input size required by the subsequent recognition model, is input to the orientation recognition unit 52. However, if the image size of the binary image input from the grayscale conversion unit 512 is equal to the input size required by the subsequent recognition model, the scaling process is not performed, and the image input from the grayscale conversion unit 512 is input directly to the orientation recognition unit 52.

[0052] Here, we will explain the methods for binarization and scaling. As described above, the binary image generation unit 511 can generate a binary image with area gradation by performing binarization by referring to surrounding pixels in the image read by the scanner. The scaling image generation unit 513 can generate a scaled image with area gradation by performing scaling by referring to surrounding pixels in the image read by the scanner.

[0053] <Processing method> Figure 13 is a flowchart of the image processing and orientation recognition processes according to one embodiment of the present invention. When image data from the HDD 3 is input to the image file format conversion device 5, the processes shown in Figure 13 begin.

[0054] In step 601 (S601), the binary image generation unit 511 checks whether the image data is a binary image. If the input image data is a binary image (Yes), the process proceeds to S603. On the other hand, if the input image data is not a binary image (i.e., a multi-level image) (No), the process proceeds to S602.

[0055] In step 602 (S602), the binary image generation unit 511 generates a binary image from the image data (i.e., the multi-level image) of S601 based on a binarization algorithm that takes area gradation into consideration. Once the processing in S602 is complete, the process proceeds to S603.

[0056] In step 603 (S603), the grayscale conversion unit 512 checks whether the image data (i.e., the binary image) from S601 or S602 is an 8-bit / pixel image. If the image data is 8 bits / pixel (Yes), the process proceeds to S605. On the other hand, if the image data is not 8 bits / pixel (No), the process proceeds to S604.

[0057] In step 604 (S604), the grayscale conversion unit 512 converts the bit depth of the image data (i.e., the binary image) to 8 bits / pixel (conversion from 0 to 0, and from 1 to 255 (grayscale conversion)). When the processing in S604 is completed, the process proceeds to S605.

[0058] In step 605 (S605), the resizing image generation unit 513 checks whether the image size of the image data (i.e., a binary image with 8 bits / pixel) in S603 or S604 matches the pre-set image size. If the image data matches the pre-set image size (Yes), proceed to S607. If the image data does not match the pre-set image size (No), proceed to S606.

[0059] In step 606 (S606), the resizing image generation unit 513 performs resizing processing on the image data based on a resizing algorithm (such as the area averaging method) that takes area gradation into account, so that the image becomes a pre-set image size. At this time, if the image data is larger than the pre-set image size, a reduction resizing process is applied. Conversely, if the image data is smaller than the pre-set image size, an enlargement resizing process is applied. When the processing in S606 is completed, the process proceeds to S607.

[0060] In step 607 (S607), the orientation discrimination unit 52 inputs the image data (pre-set image size (for example, the image size to be input to the recognition model)) into a pre-trained recognition model and performs inference. In this embodiment, the recognition model is an orientation discrimination model that outputs one of the following for the input image data: north, east, west, or south.

[0061] Figure 14 is a diagram illustrating the binarization process of image data.

[0062] For image data that is determined not to be a binary image in the S601 process in Figure 13, a binarization process based on a binarization algorithm that considers area gradation is applied. Typical binarization algorithms that consider area gradation include error diffusion.

[0063] For example, when attempting to perform processing on an input image (multi-level image) like the one shown in Figure 14(1), the binary image generation unit 511 within the image processing unit 51 generates a binary image that maintains gradation, as shown in Figure 14(3) (for reference, a binary image that does not consider area gradation is shown in Figure 14(2)).

[0064] Figure 15 is a diagram illustrating the scaling process for image data.

[0065] For image data that is determined not to be of the pre-set image size in processing S605 in Figure 13, a scaling process based on a scaling algorithm that considers area gradation is applied to make it the pre-set image size. Typical scaling algorithms that consider area gradation include the area averaging method and the Gaussian filter + bicubic method.

[0066] For example, in the case of an input image (multi-level image) like (1) in Figure 15, a binary image (2) in Figure 15 is generated by processing S601 to S604 in Figure 13, and a scaled image of a pre-set image size, like (3) in Figure 15, is generated by processing S606 in Figure 13. In the example in Figure 15, it is a reduction scaled image, but if the image size of the input image is smaller than the pre-set image size, it will be an enlargement scaled image. The generated scaled image is input to the orientation recognition unit 52, and orientation recognition inference is performed.

[0067] <Hardware Configuration> Figure 16 is a hardware configuration diagram of an image forming apparatus 100 according to one embodiment of the present invention.

[0068] As shown in Figure 16, the image forming apparatus 100 includes a controller 1010, a short-range communication circuit 1020, an engine control unit 1030, an operation panel 1040, and a network interface 1050.

[0069] Of these, the controller 1010 includes the main components of the computer: the CPU 1001, system memory (MEM-P) 1002, northbridge (NB) 1003, southbridge (SB) 1004, ASIC (Application Specific Integrated Circuit) 1005, local memory (MEM-C) 1006, HDD controller 1007, and HD 1008, with the NB 1003 and ASIC 1005 connected by an AGP (Accelerated Graphics Port) bus 1021.

[0070] Of these, the CPU 1001 is a control unit that performs overall control of the image forming apparatus 100. The NB 1003 is a bridge for connecting the CPU 1001 with the MEM-P 1002, SB 1004, and AGP bus 1021, and includes a memory controller that controls reading and writing to the MEM-P 1002, as well as a PCI (Peripheral Component Interconnect) master and an AGP target.

[0071] MEM-P1002 consists of ROM 1002a, which is a memory for storing programs and data that realize the various functions of the controller 1010, and RAM 1002b, which is used for program and data deployment and drawing during memory printing. The programs stored in RAM 1002b may be configured to be provided as installable or executable files recorded on a computer-readable recording medium such as a CD-ROM, CD-R, or DVD.

[0072] SB1004 is a bridge for connecting NB1003 to PCI devices and peripheral devices. ASIC1005 is an integrated circuit (IC) for image processing applications that has hardware elements for image processing and acts as a bridge connecting the AGP bus 1021, PCI bus 1022, HDD 1007, and MEM-C 1006, respectively. This ASIC1005 consists of a PCI target and AGP master, an arbiter (ARB) that forms the core of ASIC1005, a memory controller that controls MEM-C 1006, multiple DMACs (Direct Memory Access Controllers) that perform image data rotation using hardware logic, and a PCI unit that performs data transfer via PCI bus 1022 between the scanner unit 1031 and the printer unit 1032. Note that ASIC1005 may also be connected to a USB (Universal Serial Bus) interface or an IEEE1394 (Institute of Electrical and Electronics Engineers 1394) interface.

[0073] MEM-C1006 is local memory used as a copy image buffer and a code buffer. HD1008 is storage for storing image data, font data used during printing, and forms. HD1008 controls data reading or writing to it according to the control of CPU1001. The AGP bus 1021 is a bus interface for graphics accelerator cards proposed to speed up graphics processing, and by directly accessing MEM-P1002 with high throughput, the graphics accelerator card can be made faster.

[0074] Furthermore, the short-range communication circuit 1020 is equipped with a short-range communication circuit 1020a. The short-range communication circuit 1020 is a communication circuit such as NFC or Bluetooth.

[0075] Furthermore, the engine control unit 1030 is composed of a scanner unit 1031 and a printer unit 1032. The operation panel 1040 includes a panel display unit 1040a, such as a touch panel, which displays current settings and selection screens and accepts input from the operator, and an operation panel 1040b, which consists of a numeric keypad that accepts setting values ​​for image formation conditions such as density settings and a start key that accepts a copy start command. The controller 1010 controls the entire image forming apparatus 100, for example, controlling drawing, communication, and input from the operation panel 1040. The scanner unit 1031 or the printer unit 1032 includes an image processing section, such as error diffusion and gamma conversion.

[0076] Furthermore, the image forming apparatus 100 can be sequentially switched and selected between the document box function, copy function, printer function, and facsimile function using the application switching key on the operation panel 1040. When the document box function is selected, it enters document box mode; when the copy function is selected, it enters copy mode; when the printer function is selected, it enters printer mode; and when the facsimile mode is selected, it enters facsimile mode.

[0077] Furthermore, the network interface 1050 is an interface for data communication using a communication network. The short-range communication circuit 1020 and the network interface 1050 are electrically connected to the ASIC 1005 via the PCI bus 1022.

[0078] <Effects> Thus, in one embodiment of the present invention, a top / bottom recognition method using Deep Learning technology can achieve the same top / bottom recognition accuracy as when a multi-level image is input, even when a binary image is input. Specifically, it can improve the top / bottom recognition accuracy for documents such as photographs compared to conventional top / bottom recognition methods using OCR. Furthermore, it can achieve the same top / bottom recognition accuracy as when a multi-level image is input, even in environments where only binary images are available.

[0079] In other words, by not using OCR, it maintains its advantage over documents such as photographs without text, and furthermore, it can achieve the same orientation recognition accuracy as multi-level images even in environments where only binary images are available instead of multi-level images. In addition, even when only binary images are available, or when only binary images are available, it can improve the recognition rate of orientation recognition for binary images.

[0080] Each of the functions of the embodiments described above can be realized by one or more processing circuits. Hereinafter, "processing circuit" as used herein includes processors programmed to execute each function by software, such as processors implemented by electronic circuits, as well as devices such as ASICs (Application Specific Integrated Circuits), DSPs (digital signal processors), FPGAs (field programmable gate arrays), and conventional circuit modules designed to execute each of the functions described above.

[0081] It should be noted that the present invention is not limited to the configurations shown in the above embodiments, including combinations with other elements. These aspects can be modified without departing from the spirit of the present invention and can be appropriately determined according to their application. [Explanation of symbols]

[0082] 1 Scanner 2 Image Processing Device 3 HDD 4 plotters 5. Image file format conversion device 21 Gamma Correction Unit 22 Region detection unit 23 Data Interface Section 24-color processing / UCR section 25 Printer Correction Unit 51 Image Processing Department 52. Top and bottom identification unit 53 File Format Conversion Section 100 Image forming apparatus 511 Binary Image Generation Unit 512 Grayscale Conversion Section 513 Image scaling generation unit [Prior art documents] [Patent Documents]

[0083] [Patent Document 1] Japanese Patent Application Publication No. 05-217022

Claims

1. A binary image generation unit generates a binary image (1 bit / pixel) with area gradation from an image read by a scanner, A grayscale conversion unit that converts the aforementioned binary image (1 bit / pixel) to a binary image (8 bits / pixel), The image recognition unit outputs a classification of the aforementioned binary image (8 bits / pixel) according to a pre-trained neural network model. An image forming apparatus equipped with [a specific feature].

2. A binary image generation unit generates a binary image (1 bit / pixel) with area gradation from an image read by a scanner, A grayscale conversion unit that converts the aforementioned binary image (1 bit / pixel) to a binary image (8 bits / pixel), A scaling processing method based on a scaling algorithm that takes area gradation into consideration includes a scaling image generation unit that generates a scaled image from the binary image (8 bits / pixel), and The image recognition unit outputs a classification of the aforementioned magnified image according to a pre-trained neural network model. An image forming apparatus equipped with [a specific feature].

3. The image forming apparatus according to claim 1, wherein the binary image generation unit generates a binary image (1 bit / pixel) having area gradation by performing a binarization process that references surrounding pixels in the image read by the scanner.

4. The image forming apparatus according to claim 2, wherein the image resizing unit generates the resizing image after resizing by performing a resizing process that references surrounding pixels in the binary image (8 bits / pixel).

5. The image forming apparatus according to any one of claims 1 to 4, wherein the neural network model is a neural network model configured in multiple stages with a plurality of filters.

6. The image forming apparatus according to any one of claims 1 to 5, wherein the neural network model is a neural network model trained using grayscale images as input.

7. The image forming apparatus according to any one of claims 1 to 6, wherein the image recognition unit outputs the orientation of the generated image.

8. A method performed by an image forming apparatus, The process involves generating a binary image (1 bit / pixel) with area gradation from an image read by a scanner, and The steps include converting the aforementioned binary image (1 bit / pixel) to a binary image (8 bits / pixel), The steps include outputting a classification of the aforementioned binary image (8 bits / pixel) according to a pre-trained neural network model, and A method that includes this.

9. A method performed by an image forming apparatus, The process involves generating a binary image (1 bit / pixel) with area gradation from an image read by a scanner, and The steps include converting the aforementioned binary image (1 bit / pixel) to a binary image (8 bits / pixel), A scaling method based on a scaling algorithm that takes area gradation into consideration, comprising the steps of generating a scaled image from the binary image (8 bits / pixel), The steps include outputting a classification of the aforementioned magnified image according to a pre-trained neural network model, and A method that includes this.

10. Image forming apparatus A binary image generation unit generates a binary image (1 bit / pixel) with area gradation from an image read by a scanner. A grayscale conversion unit that converts the aforementioned binary image (1 bit / pixel) to a binary image (8 bits / pixel), A program to function as an image recognition unit that outputs a classification of the aforementioned binary image (8 bits / pixel) according to a pre-trained neural network model.

11. Image forming apparatus A binary image generation unit generates a binary image (1 bit / pixel) with area gradation from an image read by a scanner. A grayscale conversion unit that converts the aforementioned binary image (1 bit / pixel) to a binary image (8 bits / pixel), A scaling processing method based on a scaling algorithm that takes area gradation into consideration is used by a scaling image generation unit that generates a scaled image from the binary image (8 bits / pixel). A program to function as an image recognition unit that outputs a classification of the aforementioned magnified image according to a pre-trained neural network model.

Images