WO2023238726A1 - 画像処理装置、画像処理方法及び画像処理プログラム - Google Patents
画像処理装置、画像処理方法及び画像処理プログラム Download PDFInfo
- Publication number
- WO2023238726A1 WO2023238726A1 PCT/JP2023/019936 JP2023019936W WO2023238726A1 WO 2023238726 A1 WO2023238726 A1 WO 2023238726A1 JP 2023019936 W JP2023019936 W JP 2023019936W WO 2023238726 A1 WO2023238726 A1 WO 2023238726A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- image
- contour
- target image
- binary
- target
- Prior art date
Links
- 238000012545 processing Methods 0.000 title claims abstract description 157
- 238000003672 processing method Methods 0.000 title description 3
- 238000000034 method Methods 0.000 claims abstract description 105
- 230000008569 process Effects 0.000 claims abstract description 53
- 230000010339 dilation Effects 0.000 claims abstract description 9
- 239000002245 particle Substances 0.000 claims description 17
- 230000011218 segmentation Effects 0.000 abstract description 6
- 238000013136 deep learning model Methods 0.000 description 21
- 238000010586 diagram Methods 0.000 description 20
- 238000006243 chemical reaction Methods 0.000 description 17
- 238000004886 process control Methods 0.000 description 15
- 239000003550 marker Substances 0.000 description 12
- 238000002372 labelling Methods 0.000 description 9
- 238000005245 sintering Methods 0.000 description 8
- 239000011347 resin Substances 0.000 description 6
- 229920005989 resin Polymers 0.000 description 6
- 238000003860 storage Methods 0.000 description 5
- 238000012549 training Methods 0.000 description 5
- 230000002776 aggregation Effects 0.000 description 4
- 238000004220 aggregation Methods 0.000 description 4
- 238000004891 communication Methods 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 4
- 238000010191 image analysis Methods 0.000 description 4
- 238000013526 transfer learning Methods 0.000 description 4
- 230000009466 transformation Effects 0.000 description 4
- 238000004458 analytical method Methods 0.000 description 3
- 239000003086 colorant Substances 0.000 description 3
- 238000000605 extraction Methods 0.000 description 3
- 238000005520 cutting process Methods 0.000 description 2
- 238000005429 filling process Methods 0.000 description 2
- 230000006870 function Effects 0.000 description 2
- 230000000877 morphologic effect Effects 0.000 description 2
- 238000010186 staining Methods 0.000 description 2
- 238000012546 transfer Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/155—Segmentation; Edge detection involving morphological operators
Definitions
- the present disclosure relates to an image processing device, an image processing method, and an image processing program.
- Examples of area segmentation methods that identify object regions included in a target image and divide the target image according to the identified object regions include methods that use image analysis such as the watershed method, R-CNN, and Selective-Search. Examples include methods using deep learning such as the method.
- the target image is an image in which multiple non-homogeneous objects (group of objects) are densely packed together (for example, an image obtained by photographing a group of particles that contain various internal unevenness and shading and have different external shapes and sizes). In this case, it becomes difficult to realize highly accurate region division.
- the present disclosure aims to improve segmentation accuracy when segmenting an image of a non-homogeneous object group into regions.
- An image processing device includes: a binary image generation unit that generates a binary image from the target image using a first trained model that binarizes the target image into an object area and a background area and generates a binary image; a prediction unit that predicts a contour line of an object region in a binary image generated from the target image using a second learned model that predicts a contour line of an object region in a binary image; a dilated contour image generation unit that performs dilation processing on the predicted contour to generate a dilated contour image; a foreground area image generation unit that generates a foreground area image based on the binary image generated from the target image and the dilated contour image; A background pixel identified based on a binary image generated from the target image, a boundary pixel identified based on the expanded contour image, and a foreground pixel identified based on the foreground region image are used. and an area dividing unit that divides the target image into object areas using a watershed
- a second aspect of the present disclosure is the image processing device according to the first aspect, comprising: The binary image generation unit generates the binary image from the target image using trained U-Net as the first trained model.
- a third aspect of the present disclosure is the image processing device according to the second aspect, comprising: The prediction unit generates a contour image from a binary image generated from the target image using a trained U-Net as the second trained model.
- a fourth aspect of the present disclosure is the image processing device according to the third aspect, comprising:
- the foreground area image generation unit generates the foreground area image by calculating a difference between a binary image generated from the target image and the dilated contour image.
- a fifth aspect of the present disclosure is the image processing device according to the first aspect, comprising:
- the target image is an image of a particle group taken under predetermined conditions.
- a sixth aspect of the present disclosure is the image processing device according to the fifth aspect, comprising:
- the image processing apparatus includes a statistical processing unit that analyzes and performs statistical processing on an image in which the target image is divided into object regions.
- a seventh aspect of the present disclosure is the image processing device according to the sixth aspect,
- the statistical processing unit changes process conditions of a process for generating the particle group based on the result of the statistical processing.
- An eighth aspect of the present disclosure is an image processing method, comprising: Binarizing the target image into an object region and a background region, and generating a binary image from the target image using a first trained model that generates a binary image; predicting the outline of the object region in the binary image generated from the target image using a second trained model that predicts the outline of the object region in the binary image; performing dilation processing on the predicted contour to generate a dilated contour image; generating a foreground region image based on the binary image generated from the target image and the dilated contour image; A background pixel identified based on a binary image generated from the target image, a boundary pixel identified based on the expanded contour image, and a foreground pixel identified based on the foreground region image are used. and dividing the target image into object regions using the watershed method.
- a ninth aspect of the present disclosure is an image processing program, comprising: Binarizing the target image into an object region and a background region, and generating a binary image from the target image using a first trained model that generates a binary image; predicting the outline of the object region in the binary image generated from the target image using a second trained model that predicts the outline of the object region in the binary image; performing dilation processing on the predicted contour to generate a dilated contour image; generating a foreground region image based on the binary image generated from the target image and the dilated contour image; A background pixel identified based on a binary image generated from the target image, a boundary pixel identified based on the expanded contour image, and a foreground pixel identified based on the foreground region image are used. and causing the computer to execute a step of dividing the target image into object regions using the watershed method.
- FIG. 1 is a diagram showing an example of the system configuration of a process control system.
- FIG. 2 is a diagram showing an example of the hardware configuration of the image processing device.
- FIG. 3 is a diagram showing a specific example of a process control system.
- FIG. 4 is a diagram for explaining details of the target image.
- FIG. 5 is a diagram illustrating an example of the functional configuration of a region dividing section when a general watershed method is used.
- FIG. 6 is a first diagram illustrating a specific example of processing by the area dividing unit when using the general watershed method.
- FIG. 7 is a second diagram illustrating a specific example of processing by the area dividing unit when using the general watershed method.
- FIG. 8 is a diagram illustrating an example of the functional configuration of the area dividing section.
- FIG. 1 is a diagram showing an example of the system configuration of a process control system.
- FIG. 2 is a diagram showing an example of the hardware configuration of the image processing device.
- FIG. 3 is
- FIG. 9 is a first diagram showing a specific example of processing by the area dividing section.
- FIG. 10 is a second diagram showing a specific example of processing by the area dividing section.
- FIG. 11 is a diagram showing an example of the functional configuration of the statistical processing section.
- FIG. 12 is a flowchart showing the flow of process control processing.
- FIG. 13 is a flowchart showing the flow of area division processing.
- FIG. 1 is a diagram showing an example of the system configuration of a process control system.
- the process control system 100 includes a target process 110, a target image generation process, and an image processing device 130.
- the target process 110 is a particle group generation process that generates a plurality of particles (particle group).
- the particle group here includes, for example, a sintered body group produced by a sintering process, a spheroid produced by a culture process, and the like.
- the particle group generated by the target process 110 is conveyed to the target image generation process 120.
- the target image generation process 120 photographs the transported particle group under predetermined conditions to create a target image that is a target image to be processed by the image processing device 130, in which a plurality of objects (object group) are densely packed. Generate an image.
- the particle group is a sintered body group
- the sintered body group is embedded in resin to form one lump, and the formed one lump is cut.
- a target image containing a dense group of non-homogeneous objects is generated.
- the particle body is, for example, a spheroid
- the target image generation process 120 a culture solution containing three-dimensionally cultured spheroids is photographed through a microscope, so that a non-homogeneous object group is concentrated. Generate a target image.
- the target image generated by the target image generation process 120 is notified to the image processing device 130.
- An image processing program is installed in the image processing device 130, and by executing the image processing program, the image processing device 130 functions as a region dividing section 131 and a statistical processing section 132.
- the region dividing unit 131 divides the target image including the object group into regions, and notifies the statistical processing unit 132 of the divided target image for each object region.
- the region dividing unit 131 performs region division using a method that combines a method using image analysis and a method using deep learning. Specifically, by combining the watershed method, which is an example of image analysis, and U-Net, which is an example of deep learning, the images used in the watershed method (background region image, foreground region image, boundary region image) are Net, and then performs region division using the watershed method.
- the region dividing unit 131 of this embodiment can realize highly accurate region division even when the target image is an image in which a non-homogeneous object group is densely packed (details will be described later). .
- the statistical processing unit 132 analyzes the target image divided into object regions and performs statistical processing (for example, processing to total the number of objects, the size of objects, etc.).
- the statistical processing unit 132 changes the process conditions in the target process 110 based on the results of the statistical processing.
- FIG. 2 is a diagram showing an example of the hardware configuration of the image processing device.
- the image processing device 130 includes a processor 201, a memory 202, an auxiliary storage device 203, an I/F (Interface) device 204, a communication device 205, and a drive device 206. Note that each piece of hardware in the image processing device 130 is interconnected via a bus 207.
- the processor 201 includes various computing devices such as a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit).
- the processor 201 reads various programs (eg, image processing programs, etc.) onto the memory 202 and executes them.
- programs eg, image processing programs, etc.
- the memory 202 includes main storage devices such as ROM (Read Only Memory) and RAM (Random Access Memory).
- the processor 201 and the memory 202 form a so-called computer, and when the processor 201 executes various programs read onto the memory 202, the computer realizes various functions.
- the auxiliary storage device 203 stores various programs and various data used when the various programs are executed by the processor 201.
- the I/F device 204 is a connection device that connects the operating device 210 and display device 220 to the image processing device 130.
- the operating device 210 is an operating device for an operator to input various instructions to the image processing device 130
- the display device 220 is a display device for providing a display screen to the operator.
- the communication device 205 is a communication device for communicating with an external device (not shown) via a network.
- the drive device 206 is a device for setting the recording medium 230.
- the recording medium 230 here includes a medium for recording information optically, electrically, or magnetically, such as a CD-ROM, a flexible disk, or a magneto-optical disk. Further, the recording medium 230 may include a semiconductor memory or the like that electrically records information, such as a ROM or a flash memory.
- the various programs to be installed in the auxiliary storage device 203 are installed by, for example, setting the distributed recording medium 230 in the drive device 206 and reading out the various programs recorded on the recording medium 230 by the drive device 206. be done.
- various programs installed in the auxiliary storage device 203 may be installed by being downloaded from a network via the communication device 205.
- FIG. 3 is a diagram showing a specific example of a process control system.
- a specific example in this embodiment will be described for a case where the target process is a sintering process.
- the target image generation process 320 is executed.
- the target process is a sintering process
- a sintered body group collection process a resin filling process, a cutting process, a staining process, and a photographing process are executed.
- the generated sintered body group 340 is collected.
- the collected sintered body group 340 is embedded in transparent resin, and a lump 321 is generated.
- the lump 321 is cut at a plurality of locations, and a plurality of lumps 322 are generated.
- the cut surfaces of the plurality of masses 322 are dyed, and in the photographing process, the stained cut surfaces are photographed, thereby generating the target image 323.
- the target process is a sintering process
- the cut surfaces of the sintered bodies embedded in the resin are dyed and photographed in an exposed state, thereby creating an object group consisting of the cut surfaces of multiple sintered bodies.
- a target image containing the is generated.
- FIG. 4 is a diagram for explaining details of the target image.
- the target image 323 has the following characteristics.
- a group of objects consisting of cut surfaces of multiple sintered bodies are densely packed.
- the object group includes objects of various shapes and sizes (for example, object areas 401 to 405, etc.).
- the interior of each object is not uniform in color (for example, object areas 401 to 405, etc.) due to various unevenness, shading, etc. of the cut surface of the sintered body.
- Background areas (numerals 411 to 413, etc.), which are areas other than objects, are not uniform in color due to the presence or absence of a sintered body behind them.
- the region dividing unit 131 combines the watershed method and U-Net, generates an image to be used in the watershed method using U-Net, and then performs region segmentation using the watershed method. Therefore, in explaining the functional configuration of the area dividing section 131, first, ⁇ Functional configuration of the area division part when using the general watershed method, will be explained, followed by -Functional configuration of the area dividing unit 131 according to the present embodiment, which combines the watershed method and U-Net; The characteristics of the area dividing unit 131 according to this embodiment will be clarified by explaining.
- FIG. 5 is a diagram illustrating an example of the functional configuration of a region dividing section when a general watershed method is used.
- FIGS. 6 and 7 are first and second diagrams showing specific examples of processing by the area division unit when using the general watershed method.
- a region dividing unit 500 using a general watershed method includes a threshold processing unit 501, a morphology conversion processing unit 502, a distance conversion processing unit 503, a difference processing unit 504, a labeling processing unit 505, and a watershed method. It has a processing section 506.
- the threshold processing unit 501 performs threshold processing on the target image (for example, the target image 323 (see FIGS. 3, 4, and 6)) under the adjusted threshold parameter, and blackens pixels that are less than the threshold parameter. Next, a binary image is generated by converting pixels whose color is equal to or greater than a threshold parameter to white. The binary image generated by the threshold processing unit 501 is notified to the morphology conversion processing unit 502.
- the morphological conversion processing unit 502 performs morphological conversion processing on the binary image generated by the threshold processing unit 501 under the adjusted conversion parameters to convert the background area image (for example, the background area image 511 (Fig. 6)) and identify the background pixels.
- the background area image (for example, background area image 511 (see FIG. 6)) generated by the morphology conversion processing unit 502 is notified to the distance conversion processing unit 503 and the difference processing unit 504.
- the distance conversion processing unit 503 converts the pixel value of each pixel other than the background area in the background area image generated by the morphology conversion processing unit 502 into a pixel according to the distance from the background pixel based on the adjusted conversion parameter. Convert to value. Thereby, the distance conversion processing unit 503 generates a distance image (for example, the distance image 512 (see FIG. 6)).
- the distance conversion processing unit 503 identifies pixels whose value is equal to or greater than the adjusted threshold parameter in the distance image (for example, the distance image 512 (see FIG. 6)) as foreground pixels. Furthermore, the distance conversion processing unit 503 generates a foreground area image (for example, the foreground area image 513 (see FIG. 6)) in which the identified foreground pixels are white. The foreground area image (for example, foreground area image 513 (see FIG. 6)) generated by the distance conversion processing unit 503 is notified to the difference processing unit 504.
- the difference processing unit 504 uses a background area image (for example, background area image 511 (see FIG. 6)) generated by the morphology conversion processing unit 502 and a foreground area image (for example, the foreground area image) generated by the distance conversion processing unit 503. image 513 (see FIG. 6)). Thereby, the difference processing unit 504 generates a boundary area image (eg, boundary area image 514 (see FIG. 7)) and identifies boundary pixels.
- a background area image for example, background area image 511 (see FIG. 6)
- a foreground area image for example, the foreground area image generated by the distance conversion processing unit 503. image 513 (see FIG. 6)
- the difference processing unit 504 generates a boundary area image (eg, boundary area image 514 (see FIG. 7)) and identifies boundary pixels.
- the labeling processing unit 505 - Background pixels identified based on a background area image (for example, background area image 511 (see FIG. 6)); - Foreground pixels identified based on a foreground region image (for example, foreground region image 513 (see FIG. 6)); - Boundary pixels identified based on a boundary area image (for example, boundary area image 514 (see FIG. 7));
- a marker image for example, marker image 521 (see FIG. 7)
- the marker image (for example, the marker image 521 (see FIG. 7)) generated by the labeling processing unit 505 is notified to the watershed method processing unit 506.
- the watershed method processing unit 506 acquires a target image (for example, the target image 323 (see FIGS. 3, 4, and 6)) and a marker image (for example, the marker image 521 (see FIG. 7)), and uses the watershed method to , divide the target image into object regions. As a result, the watershed method processing unit 506 generates and outputs a region divided image (for example, a region divided image 522 (see FIG. 7) that is color-coded so that the regions of each object can be identified). Alternatively, the watershed method processing unit 506 generates and outputs a region segmented image (for example, a region segmented image 523 (see FIG. 7) that is color-coded so that the outline of each object can be seen).
- a target image for example, the target image 323 (see FIGS. 3, 4, and 6)
- a marker image for example, the marker image 521 (see FIG. 7)
- threshold processing section 501 As described above, it is necessary to appropriately adjust the threshold parameters or transformation parameters in the threshold processing section 501, morphology transformation processing section 502, and distance transformation processing section 503 of the region division section 500.
- the target image 323 an image with a dense group of non-homogeneous objects
- an object area that should originally be divided into one object area 402 may be divided into two object areas. can occur.
- object area 403' in FIG. 7 a situation may occur in which an object area that should originally be divided into three object areas 403 to 405 is divided into one object area.
- FIG. 8 is a diagram illustrating an example of the functional configuration of the area dividing section.
- FIGS. 9 and 10 are first and second diagrams showing specific examples of processing by the area dividing unit.
- the region dividing unit 131 includes a U-Net mask processing unit 801, a U-Net contour prediction unit 802, an expanded contour image generation unit 803, a difference processing unit 804, a labeling processing unit 505, and a watershed method. It has a processing section 506.
- the U-Net mask processing unit 801 is an example of a binary image generation unit.
- the U-Net mask processing unit 801 uses a trained deep learning model (first trained model) that performs mask processing on the input image (predicts the object and background and generates a binary image). example), which has a deep learning model for the target image.
- the U-Net mask processing unit 801 has a network structure of U-Net, and in this embodiment, ⁇ Input image: An image containing a group of objects made up of cut surfaces of multiple sintered bodies by photographing the cut surfaces of a sintered body embedded in resin in a dyed and exposed state; ⁇ Correct image: An image in which the input image is processed so that the object group is white and the background is black, It has a trained deep learning model (that is, a trained U-Net) that has undergone transfer learning using training data including . Note that the trained deep learning model can be generated by performing transfer learning on an existing trained deep learning model.
- ⁇ Input image photographic image obtained by photographing a car
- ⁇ Correct image In the input image, the image where the car is white and the background is black
- a trained U-Net can be generated.
- the U-Net mask processing unit 801 By inputting a target image (for example, the target image 323 (see FIGS. 3, 4, and 9)), the U-Net mask processing unit 801 generates a binary image (for example, the binary image 811 (see FIG. 9)). )). Thereby, the U-Net mask processing unit 801 identifies background pixels.
- a target image for example, the target image 323 (see FIGS. 3, 4, and 9)
- a binary image for example, the binary image 811 (see FIG. 9)
- the binary image generated by the U-Net mask processing unit 801 (for example, the binary image 811 (see FIG. 9)) is notified to the U-Net contour prediction unit 802 and the difference processing unit 804.
- the U-Net contour prediction unit 802 is an example of a prediction unit.
- the U-Net contour prediction unit 802 has a trained deep learning model (an example of a second trained model) that predicts a contour from an input binary image.
- the U-Net contour prediction unit 802 has a network structure of U-Net, and in this embodiment, ⁇ Input image: An image in which objects are colored white and the background is colored black, ⁇ Correct image: An image in which the outline of each object in the object group is made white, and objects other than the outline and the background are made black, It has a trained deep learning model (that is, a trained U-Net) trained using training data including the following.
- the U-Net contour prediction unit 802 predicts a contour and outputs a contour image. Note that the contour image output by the U-Net contour prediction unit 802 is notified to the dilated contour image generation unit 803.
- the dilated contour image generation unit 803 expands the contour in the contour image output by the U-Net contour prediction unit 802 to create a dilated contour image (for example, dilated contour image 812 (see FIG. 9). )). Thereby, the dilated contour image generation unit 803 identifies boundary pixels.
- the dilated contour image generated by the dilated contour image generation unit 803 (for example, the dilated contour image 812 (see FIG. 9)) is notified to the difference processing unit 804.
- the difference processing unit 804 is an example of a foreground area image generation unit.
- the difference processing unit 804 generates a binary image (for example, a binary image 811 (see FIG. 9)) generated by the U-Net mask processing unit 801 and an expanded contour image generated by the expanded contour image generation unit 803. (For example, the dilated contour image 812 (see FIG. 9)).
- the difference processing unit 804 generates a foreground area image (for example, the foreground area image 813 (see FIG. 9)) and identifies foreground pixels.
- the labeling processing unit 505 - Background pixels identified based on a binary image (for example, binary image 811 (see FIG. 9)); - Boundary pixels identified based on the dilated contour image (for example, dilated contour image 812 (see FIG. 9)); Foreground pixels identified based on a foreground region image (for example, foreground region image 813 (see FIG. 9)); By labeling each, a marker image (for example, marker image 821 (see FIG. 10)) is generated. The marker image (for example, marker image 821 (see FIG. 10)) generated by the labeling processing unit 505 is notified to the watershed method processing unit 506.
- a marker image for example, marker image 821 (see FIG. 10)
- the watershed method processing unit 506 acquires a target image (for example, the target image 323 (see FIGS. 3, 4, and 9)) and a marker image (for example, the marker image 821 (see FIG. 10)), and uses the watershed method to , divide the target image into object regions. As a result, the watershed method processing unit 506 generates and outputs a region divided image (for example, a region divided image 822 (see FIG. 10) that is color-coded so that the regions of each object can be identified). Alternatively, the watershed method processing unit 506 generates and outputs a region segmented image (for example, a region segmented image 823 (see FIG. 10) colored so that the outline of each object can be seen).
- a target image for example, the target image 323 (see FIGS. 3, 4, and 9)
- a marker image for example, the marker image 821 (see FIG. 10)
- region division processing by the region division section 131 according to this embodiment which combines the watershed method and U-Net
- - In the area division process by the area division unit 500, various parameters needed to be adjusted in order to generate a background area image, a foreground area image, and a boundary area image for specifying background pixels, foreground pixels, and boundary pixels.
- region division processing by the region division unit 131 according to the present embodiment since a trained deep learning model is used, adjustment of various parameters is not necessary.
- the contour line can be predicted with high accuracy.
- the non-uniform colors inside the object are removed, so unlike when predicting the contour line directly from the target image, this is due to the influence of the non-uniform colors inside the object. This is because the contour line can be predicted without being affected.
- the region dividing unit 131 can realize highly accurate region division even when the target image is an image in which a non-homogeneous object group is densely packed.
- FIG. 11 is a diagram showing an example of the functional configuration of the statistical processing section.
- the statistical processing unit 132 includes an object extraction unit 1101, an analysis unit 1102, an aggregation unit 1103, and a condition change unit 1104.
- the object extracting unit 1101 obtains a region divided image (for example, the region divided image 822 (see FIG. 10) or the region divided image 823 (see FIGS. 10 and 11)) from the region dividing unit 131.
- the object extraction unit 1101 extracts an image 1121 or the like in object units based on the obtained region-divided image (for example, the region-divided image 822 (see FIG. 10) or the region-divided image 823 (see FIGS. 10 and 11)). generate.
- the analysis unit 1102 analyzes the object-based image 1121 generated by the object extraction unit 1101, and identifies, for example, the type and size of the object.
- the aggregation unit 1103 aggregates the types, sizes, etc. of objects identified by the analysis unit 1102, and outputs the aggregation results.
- the condition changing unit 1104 changes the process conditions for the sintered body based on the tally results output by the tallying unit 1103, and outputs the changed process conditions.
- FIG. 12 is a flowchart showing the flow of process control processing.
- step S1201 the sintering process 310 is executed to generate a sintered body group 340.
- step S1202 the target image generation process 320 is executed to generate a target image 323 that includes a non-homogeneous object group.
- step S1203 the image processing device 130 performs region division processing to divide the target image 323 into object regions. Note that details of the area division process will be described later.
- step S1204 the image processing device 130 analyzes the region segmented image 822 or the region segmented image 823 generated by performing the region segmentation process, performs statistical processing, and outputs a total result.
- step S1205 the image processing device 130 changes the process conditions based on the tally result, and outputs the changed process conditions.
- step S1206 the sintering process 310 is executed under the changed process conditions to generate a sintered body group 340.
- FIG. 13 is a flowchart showing the flow of area division processing.
- step S1301 the image processing device 130 acquires the target image 323.
- step S1302 the image processing device 130 generates a binary image 811 from the target image 323 using a trained deep learning model whose network structure is U-Net, and identifies background pixels.
- step S1303 the image processing device 130 uses a trained deep learning model whose network structure is U-Net to predict a contour from the binary image and generate a contour image.
- step S1304 the image processing device 130 performs dilation processing on the contour image to generate a dilated contour image 812 to identify boundary pixels.
- step S1305 the image processing device 130 generates a foreground region image 813 based on the difference between the binary image 811 and the dilated contour image 812, and identifies foreground pixels.
- step S1306 the image processing device 130 uses the identified background pixels, boundary pixels, and foreground pixels to divide the target image 323 into object regions using the watershed method.
- a deep learning model U-Net
- - Perform dilation processing on the predicted contour to generate a dilated contour image.
- the target image is divided into object regions using the watershed method.
- a contour can be predicted with high accuracy using less learning data compared to a case where a contour is directly predicted from a target image.
- the network structure of the deep learning model is not limited to U-Net.
- any other deep learning model can be applied as long as it has a convolutional layer and has a network structure that can input an image and output an image.
- a deep learning model that can output the outline of each object is preferable.
- the particle group generated by the target process 110 is not limited to the sintered body group, but may be any other particle group.
- other arbitrary particle groups include spheroids.
- the region dividing unit 131 outputs the region divided image 822 or the region divided image 823, but the image outputted by the region dividing unit 131 is the region divided image 822 or It is not limited to the divided image 823, and may be another image.
- the statistical processing unit 132 outputs a tally result in which the type of each object and the size of each object are tallied.
- the items to be included are not limited to these.
- Process control system 130 Image processing device 131: Area division unit 132: Statistical processing unit 323: Target image 505: Labeling processing unit 506: Watershed method processing unit 801: U-Net mask processing unit 802: U-Net contour line Prediction unit 803: Dilated contour image generation unit 804: Difference processing unit
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
Description
対象画像をオブジェクト領域と背景領域とに2値化し、2値画像を生成する第1の学習済みモデルを用いて、対象画像から2値画像を生成する2値画像生成部と、
2値画像においてオブジェクト領域の輪郭線を予測する第2の学習済みモデルを用いて、前記対象画像から生成された2値画像において、オブジェクト領域の輪郭線を予測する予測部と、
前記予測した輪郭線に対して膨張処理を行い、膨張輪郭線画像を生成する膨張輪郭線画像生成部と、
前記対象画像から生成された2値画像と前記膨張輪郭線画像とに基づいて、前景領域画像を生成する前景領域画像生成部と、
前記対象画像から生成された2値画像に基づいて特定される背景画素と、前記膨張輪郭線画像に基づいて特定される境界画素と、前記前景領域画像に基づいて特定される前景画素とを用いて、分水嶺法により前記対象画像をオブジェクト領域ごとに分割する領域分割部とを有する。
前記2値画像生成部は、前記第1の学習済みモデルとして、学習済みのU-Netを用いて、前記対象画像から前記2値画像を生成する。
前記予測部は、前記第2の学習済みモデルとして、学習済みのU-Netを用いて、前記対象画像から生成された2値画像から輪郭線画像を生成する。
前記前景領域画像生成部は、前記対象画像から生成された2値画像と、前記膨張輪郭線画像との差分を算出することで、前記前景領域画像を生成する。
前記対象画像は、所定の条件のもとで粒子体群を撮影した画像である。
前記対象画像がオブジェクト領域ごとに分割された画像を解析し、統計処理する統計処理部を有する。
前記統計処理部は、統計処理の結果に基づいて、前記粒子体群を生成するプロセスのプロセス条件を変更する。
対象画像をオブジェクト領域と背景領域とに2値化し、2値画像を生成する第1の学習済みモデルを用いて、対象画像から2値画像を生成する工程と、
2値画像においてオブジェクト領域の輪郭線を予測する第2の学習済みモデルを用いて、前記対象画像から生成された2値画像において、オブジェクト領域の輪郭線を予測する工程と、
前記予測した輪郭線に対して膨張処理を行い、膨張輪郭線画像を生成する工程と、
前記対象画像から生成された2値画像と前記膨張輪郭線画像とに基づいて、前景領域画像を生成する工程と、
前記対象画像から生成された2値画像に基づいて特定される背景画素と、前記膨張輪郭線画像に基づいて特定される境界画素と、前記前景領域画像に基づいて特定される前景画素とを用いて、分水嶺法により前記対象画像をオブジェクト領域ごとに分割する工程とをコンピュータが実行する。
対象画像をオブジェクト領域と背景領域とに2値化し、2値画像を生成する第1の学習済みモデルを用いて、対象画像から2値画像を生成する工程と、
2値画像においてオブジェクト領域の輪郭線を予測する第2の学習済みモデルを用いて、前記対象画像から生成された2値画像において、オブジェクト領域の輪郭線を予測する工程と、
前記予測した輪郭線に対して膨張処理を行い、膨張輪郭線画像を生成する工程と、
前記対象画像から生成された2値画像と前記膨張輪郭線画像とに基づいて、前景領域画像を生成する工程と、
前記対象画像から生成された2値画像に基づいて特定される背景画素と、前記膨張輪郭線画像に基づいて特定される境界画素と、前記前景領域画像に基づいて特定される前景画素とを用いて、分水嶺法により前記対象画像をオブジェクト領域ごとに分割する工程とをコンピュータに実行させる。
<プロセス制御システムのシステム構成>
はじめに、第1の実施形態に係る画像処理装置を備えるプロセス制御システム全体のシステム構成について説明する。図1は、プロセス制御システムのシステム構成の一例を示す図である。
次に、画像処理装置130のハードウェア構成について説明する。図2は、画像処理装置のハードウェア構成の一例を示す図である。
次に、対象プロセスが焼結プロセスである場合のプロセス制御システム100の具体例について説明する。図3は、プロセス制御システムの具体例を示す図である。以降、本実施形態における具体例の説明は、対象プロセスが焼結プロセスである場合について行う。
次に、対象画像生成プロセス320により生成された対象画像323の詳細について説明する。図4は、対象画像の詳細を説明するための図である。
・複数の焼結体の切断面からなるオブジェクト群が密集している。
・当該オブジェクト群には、様々な形状及び大きさのオブジェクトが含まれる(例えば、オブジェクト領域401~405等)。
・各オブジェクトの内部は、焼結体の切断面の様々な凹凸や濃淡等に起因して、色が均一でない(例えば、オブジェクト領域401~405等)。
・オブジェクト以外の領域である背景領域(符号411~413等)は、背後の焼結体の有無に起因して、色が均一ではない。
次に、対象画像323をオブジェクト領域ごとに分割する領域分割部131の機能構成について説明する。上述したように、領域分割部131は、分水嶺法とU-Netとを組み合わせ、分水嶺法に用いる画像をU-Netを用いて生成したうえで、分水嶺法により領域分割を行う。そこで、以下、領域分割部131の機能構成を説明するにあたっては、はじめに、
・一般的な分水嶺法を用いた場合の領域分割部の機能構成、
について説明し、続いて、
・分水嶺法とU-Netとを組み合わせた本実施形態に係る領域分割部131の機能構成、
を説明することで、本実施形態に係る領域分割部131の特徴を明確にする。
はじめに、一般的な分水嶺法を用いた場合の領域分割部の機能構成について、図6及び図7を参照しながら、図5を用いて説明する。図5は、一般的な分水嶺法を用いた場合の領域分割部の機能構成の一例を示す図である。また、図6、図7は、一般的な分水嶺法を用いた場合の領域分割部による処理の具体例を示す第1及び第2の図である。
・背景領域画像(例えば、背景領域画像511(図6参照))に基づいて特定された背景画素、
・前景領域画像(例えば、前景領域画像513(図6参照))に基づいて特定された前景画素、
・境界領域画像(例えば、境界領域画像514(図7参照))に基づいて特定された境界画素、
それぞれにラベリングを行うことで、マーカ画像(例えば、マーカ画像521(図7参照))を生成する。ラベリング処理部505により生成されたマーカ画像(例えば、マーカ画像521(図7参照))は、分水嶺法処理部506に通知される。
次に、分水嶺法とU-Netとを組み合わせた本実施形態に係る領域分割部131の機能構成について、図9、図10を参照しながら図8を用いて説明する。図8は、領域分割部の機能構成の一例を示す図である。また、図9、図10は、領域分割部による処理の具体例を示す第1及び第2の図である。
・入力画像:樹脂に埋め込まれた焼結体の切断面が染色され露出された状態で撮影されることで、複数の焼結体の切断面からなるオブジェクト群が含まれる画像、
・正解画像:入力画像において、オブジェクト群を白色、背景を黒色に加工した画像、
を含む学習用データを用いて転移学習された学習済みの深層学習モデル(つまり、学習済みのU-Net)を有する。なお、当該学習済みの深層学習モデルは、既存の学習済みの深層学習モデルを転移学習させることで生成することができる。例えば、
・入力画像:車を撮影することで得られた写真画像、
・正解画像:入力画像において、車を白色、背景を黒色とした画像、
を学習用データとして学習された既存の学習済みの深層学習モデルを、更に、対象画像用に転移学習させることで、対象画像用に転移学習された学習済みの深層学習モデル(つまり、転移学習された学習済みのU-Net)を生成することができる。既存の学習済みの深層学習モデルを転移学習に用いることで、新たに必要な学習用データの数を削減することができるとともに、学習にかかる時間を削減することができる。
・入力画像:オブジェクト群を白色、背景を黒色に加工した画像、
・正解画像:オブジェクト群の各オブジェクトの輪郭線を白色、輪郭線以外のオブジェクト及び背景を黒色に加工した画像、
を含む学習用データを用いて学習された学習済みの深層学習モデル(つまり、学習済みのU-Net)を有する。
・2値画像(例えば、2値画像811(図9参照))に基づいて特定された背景画素、
・膨張輪郭線画像(例えば、膨張輪郭線画像812(図9参照))に基づいて特定された境界画素、
・前景領域画像(例えば、前景領域画像813(図9参照))に基づいて特定された前景画素、
それぞれにラベリングを行うことで、マーカ画像(例えば、マーカ画像821(図10参照))を生成する。ラベリング処理部505により生成されたマーカ画像(例えば、マーカ画像821(図10参照))は、分水嶺法処理部506に通知される。
・領域分割部500による領域分割処理では、背景画素、前景画素、境界画素を特定するための背景領域画像、前景領域画像、境界領域画像を生成するにあたり、各種パラメータの調整が必要であった。これに対して、本実施形態に係る領域分割部131による領域分割処理の場合、学習済みの深層学習モデルを用いるため、各種パラメータの調整が不要である。
・境界画素を特定するための輪郭線画像を出力するにあたり、対象画像から直接的に輪郭線を予測する代わりに、対象画像からオブジェクト領域と背景領域とを予測し、2値画像を生成したうえで輪郭線を予測し、輪郭線画像を出力する。これにより、本実施形態に係る領域分割部131の場合、輪郭線を高精度に予測することができる。これは、2値画像の場合、オブジェクトの内部の不均一な色が削除されているため、対象画像から直接的に輪郭線を予測する場合とは異なり、オブジェクトの内部の不均一な色の影響を受けることなく輪郭線を予測することができるからである。また、対象画像から、オブジェクト領域と背景領域を予測し2値画像を生成する際も、対象画像から輪郭線を予測する場合と比較して、オブジェクトの内部の不均一な色の影響を受けることが少なくて済むからである。つまり、2種類の学習済みの深層学習モデルを組み合わせて、2段階で輪郭線を予測する構成とすることで、均質でないオブジェクト群の影響を低減できるため、結果的に、より少ない学習用データで、輪郭線を高精度に予測することができる。
次に、統計処理部132の機能構成について説明する。図11は、統計処理部の機能構成の一例を示す図である。図11に示すように、統計処理部132は、オブジェクト抽出部1101、解析部1102、集計部1103、条件変更部1104を有する。
次に、プロセス制御システム100によるプロセス制御処理の流れについて説明する。図12は、プロセス制御処理の流れを示すフローチャートである。
次に、領域分割処理(ステップS1203)の詳細について説明する。図13は、領域分割処理の流れを示すフローチャートである。
以上の説明から明らかなように、第1の実施形態に係る画像処理装置130は、
・対象画像をオブジェクト領域と背景領域とに2値化し、2値画像を生成する深層学習モデル(ネットワーク構造=U-Net)を用いて、対象画像から2値画像を生成する。
・2値画像においてオブジェクト領域の輪郭線を予測する深層学習モデル(ネットワーク構造=U-Net)を用いて、対象画像から生成された2値画像において、オブジェクト領域の輪郭線を予測する。
・予測した輪郭線に対して膨張処理を行い、膨張輪郭線画像を生成する。
・対象画像から生成された2値画像と、膨張輪郭線画像との差分から、前景領域画像を生成する。
・対象画像から生成された2値画像に基づいて特定される背景画素と、膨張輪郭線画像に基づいて特定される境界画素と、前景領域画像に基づいて特定される前景画素とを用いて、分水嶺法により対象画像をオブジェクト領域ごとに分割する。
上記第1の実施形態では、深層学習モデルのネットワーク構造として、U-Netを用いる場合について説明した。しかしながら、深層学習モデルのネットワーク構造は、U-Netに限定されない。例えば、畳み込み層を有し、画像を入力することで、画像を出力可能なネットワーク構造であれば、他の任意の深層学習モデルが適用可能である。ただし、オブジェクト領域の輪郭線の予測をより適切に行うことができるよう、個々のオブジェクトの輪郭を出力可能な深層学習モデルであることが好ましい。
130 :画像処理装置
131 :領域分割部
132 :統計処理部
323 :対象画像
505 :ラベリング処理部
506 :分水嶺法処理部
801 :U-Netマスク処理部
802 :U-Net輪郭線予測部
803 :膨張輪郭線画像生成部
804 :差分処理部
Claims (9)
- 対象画像をオブジェクト領域と背景領域とに2値化し、2値画像を生成する第1の学習済みモデルを用いて、対象画像から2値画像を生成する2値画像生成部と、
2値画像においてオブジェクト領域の輪郭線を予測する第2の学習済みモデルを用いて、前記対象画像から生成された2値画像において、オブジェクト領域の輪郭線を予測する予測部と、
前記予測した輪郭線に対して膨張処理を行い、膨張輪郭線画像を生成する膨張輪郭線画像生成部と、
前記対象画像から生成された2値画像と前記膨張輪郭線画像とに基づいて、前景領域画像を生成する前景領域画像生成部と、
前記対象画像から生成された2値画像に基づいて特定される背景画素と、前記膨張輪郭線画像に基づいて特定される境界画素と、前記前景領域画像に基づいて特定される前景画素とを用いて、分水嶺法により前記対象画像をオブジェクト領域ごとに分割する領域分割部と
を有する画像処理装置。 - 前記2値画像生成部は、前記第1の学習済みモデルとして、学習済みのU-Netを用いて、前記対象画像から前記2値画像を生成する、請求項1に記載の画像処理装置。
- 前記予測部は、前記第2の学習済みモデルとして、学習済みのU-Netを用いて、前記対象画像から生成された2値画像から輪郭線画像を生成する、請求項2に記載の画像処理装置。
- 前記前景領域画像生成部は、前記対象画像から生成された2値画像と、前記膨張輪郭線画像との差分を算出することで、前記前景領域画像を生成する、請求項3に記載の画像処理装置。
- 前記対象画像は、所定の条件のもとで粒子体群を撮影した画像である、請求項1乃至4のいずれか1項に記載の画像処理装置。
- 前記対象画像がオブジェクト領域ごとに分割された画像を解析し、統計処理する統計処理部を有する、請求項5に記載の画像処理装置。
- 前記統計処理部は、統計処理の結果に基づいて、前記粒子体群を生成するプロセスのプロセス条件を変更する、請求項6に記載の画像処理装置。
- 対象画像をオブジェクト領域と背景領域とに2値化し、2値画像を生成する第1の学習済みモデルを用いて、対象画像から2値画像を生成する工程と、
2値画像においてオブジェクト領域の輪郭線を予測する第2の学習済みモデルを用いて、前記対象画像から生成された2値画像において、オブジェクト領域の輪郭線を予測する工程と、
前記予測した輪郭線に対して膨張処理を行い、膨張輪郭線画像を生成する工程と、
前記対象画像から生成された2値画像と前記膨張輪郭線画像とに基づいて、前景領域画像を生成する工程と、
前記対象画像から生成された2値画像に基づいて特定される背景画素と、前記膨張輪郭線画像に基づいて特定される境界画素と、前記前景領域画像に基づいて特定される前景画素とを用いて、分水嶺法により前記対象画像をオブジェクト領域ごとに分割する工程と
をコンピュータが実行する画像処理方法。 - 対象画像をオブジェクト領域と背景領域とに2値化し、2値画像を生成する第1の学習済みモデルを用いて、対象画像から2値画像を生成する工程と、
2値画像においてオブジェクト領域の輪郭線を予測する第2の学習済みモデルを用いて、前記対象画像から生成された2値画像において、オブジェクト領域の輪郭線を予測する工程と、
前記予測した輪郭線に対して膨張処理を行い、膨張輪郭線画像を生成する工程と、
前記対象画像から生成された2値画像と前記膨張輪郭線画像とに基づいて、前景領域画像を生成する工程と、
前記対象画像から生成された2値画像に基づいて特定される背景画素と、前記膨張輪郭線画像に基づいて特定される境界画素と、前記前景領域画像に基づいて特定される前景画素とを用いて、分水嶺法により前記対象画像をオブジェクト領域ごとに分割する工程と
をコンピュータに実行させるための画像処理プログラム。
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202380042720.1A CN119256334A (zh) | 2022-06-06 | 2023-05-29 | 图像处理装置、图像处理方法及图像处理程序 |
JP2024526384A JPWO2023238726A1 (ja) | 2022-06-06 | 2023-05-29 |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2022-091556 | 2022-06-06 | ||
JP2022091556 | 2022-06-06 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2023238726A1 true WO2023238726A1 (ja) | 2023-12-14 |
Family
ID=89118212
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2023/019936 WO2023238726A1 (ja) | 2022-06-06 | 2023-05-29 | 画像処理装置、画像処理方法及び画像処理プログラム |
Country Status (3)
Country | Link |
---|---|
JP (1) | JPWO2023238726A1 (ja) |
CN (1) | CN119256334A (ja) |
WO (1) | WO2023238726A1 (ja) |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2018116391A (ja) * | 2017-01-17 | 2018-07-26 | みずほ情報総研株式会社 | 画像処理システム、画像処理方法及び画像処理プログラム |
KR20200095873A (ko) * | 2019-02-01 | 2020-08-11 | 한국전자통신연구원 | 인물 영역 추출 방법, 이를 이용하는 영상 처리 장치 및 인물 영역 추출 시스템 |
US20210192715A1 (en) * | 2019-05-01 | 2021-06-24 | Inspect Technologies Ltd | Automated grains inspection |
JP2021120815A (ja) | 2020-01-30 | 2021-08-19 | 富士通株式会社 | 推定処理プログラム、推定処理方法、及び情報処理装置 |
US11288535B1 (en) * | 2021-05-12 | 2022-03-29 | Zhejiang University | Method and system for identifying field boundary |
CN114494946A (zh) * | 2021-12-30 | 2022-05-13 | 青岛浩海网络科技股份有限公司 | 一种用于森林防火监控中的防太阳误报方法 |
JP2022091556A (ja) | 2020-12-09 | 2022-06-21 | 株式会社日立ハイテクサイエンス | 赤外レーザ顕微鏡 |
-
2023
- 2023-05-29 JP JP2024526384A patent/JPWO2023238726A1/ja active Pending
- 2023-05-29 WO PCT/JP2023/019936 patent/WO2023238726A1/ja active Application Filing
- 2023-05-29 CN CN202380042720.1A patent/CN119256334A/zh active Pending
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2018116391A (ja) * | 2017-01-17 | 2018-07-26 | みずほ情報総研株式会社 | 画像処理システム、画像処理方法及び画像処理プログラム |
KR20200095873A (ko) * | 2019-02-01 | 2020-08-11 | 한국전자통신연구원 | 인물 영역 추출 방법, 이를 이용하는 영상 처리 장치 및 인물 영역 추출 시스템 |
US20210192715A1 (en) * | 2019-05-01 | 2021-06-24 | Inspect Technologies Ltd | Automated grains inspection |
JP2021120815A (ja) | 2020-01-30 | 2021-08-19 | 富士通株式会社 | 推定処理プログラム、推定処理方法、及び情報処理装置 |
JP2022091556A (ja) | 2020-12-09 | 2022-06-21 | 株式会社日立ハイテクサイエンス | 赤外レーザ顕微鏡 |
US11288535B1 (en) * | 2021-05-12 | 2022-03-29 | Zhejiang University | Method and system for identifying field boundary |
CN114494946A (zh) * | 2021-12-30 | 2022-05-13 | 青岛浩海网络科技股份有限公司 | 一种用于森林防火监控中的防太阳误报方法 |
Non-Patent Citations (3)
Title |
---|
ANONYMOUS: "[Python] Watershed algorithm and object segmentation with OpenCV", CODEGRAFFITI, 26 March 2020 (2020-03-26), pages 1 - 22, XP093114707, Retrieved from the Internet <URL: https://code-graffiti.com/opencv-watershed-algorithm-in-python> [retrieved on 20231221] * |
IGARASHI, RIKI; KOIZUMI, NORIHIRO; NISHIYAMA, YU; TOMITA, KYOHEI; SHIGENARI, YUKA; SHOJI, SUNAO: "1A1-B09 Matching axial images of magnetic resonance imaging and transrectal ultrasound based on deep learning", THE JAPANESE SOCIETY OF MECHANICAL ENGINEERS, JAPAN SOCIETY OF MECHANICAL ENGINEERS; ROBOTICS AND MECHATRONICS DIVISION, JP, 1 January 2019 (2019-01-01) - 8 June 2019 (2019-06-08), JP , pages 141 - 141-B09-4, XP009551216, ISSN: 2424-3124, DOI: 10.1299/jsmermd.2019.1A1-B09 * |
WATANABE, HIDENORI: "Field heat map based on the perspective of an expert in growth diagnosis", IEICE TECHNICAL REPORT, SP, IEICE, JP, vol. 119, no. 441 (SP2019-115), 3 March 2020 (2020-03-03), JP, pages 379 - 384, XP009551213 * |
Also Published As
Publication number | Publication date |
---|---|
JPWO2023238726A1 (ja) | 2023-12-14 |
CN119256334A (zh) | 2025-01-03 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109961049B (zh) | 一种复杂场景下香烟品牌识别方法 | |
CN110176054B (zh) | 用于训练神经网络模型的合成图像的生成 | |
JP5174040B2 (ja) | 画像の構成要素と背景とを区別するためにコンピュータで実行される方法および画像の構成要素と背景とを区別するためのシステム | |
CN111210498B (zh) | 降低多边形网格的细节水平以减少被渲染几何的复杂度 | |
CN111145188B (zh) | 一种基于ResNet与UNet模型的图像分割方法 | |
CN110570352B (zh) | 图像标注方法、装置、系统及细胞标注方法 | |
CN112258610B (zh) | 图像标注方法、装置、存储介质及电子设备 | |
WO2023050651A1 (zh) | 图像语义分割方法、装置、设备及存储介质 | |
US12112533B2 (en) | Method and apparatus for data calculation in neural network model, and image processing method and apparatus | |
CN114723636A (zh) | 基于多特征融合的模型生成方法、装置、设备及存储介质 | |
CN116311412A (zh) | 一种融合3d注意力机制和空洞卷积的口罩佩戴检测方法 | |
US20170011552A1 (en) | 3D Model Enhancement | |
US11954799B2 (en) | Computer-implemented method for generating a 3-dimensional wireframe model of an object comprising a plurality of parts | |
Krenmayr et al. | DilatedToothSegNet: tooth segmentation network on 3d dental meshes through increasing receptive vision | |
CN107689041A (zh) | 一种用于图像相似度比对的方法、装置及存储介质 | |
CN115270184A (zh) | 视频脱敏、车辆的视频脱敏方法、车载处理系统 | |
CN113506305B (zh) | 三维点云数据的图像增强方法、语义分割方法及装置 | |
WO2023238726A1 (ja) | 画像処理装置、画像処理方法及び画像処理プログラム | |
JP2015158737A (ja) | 画像処理装置及び画像処理方法 | |
CN116342800B (zh) | 一种多模态位姿优化的语义三维重建方法及系统 | |
CN117011158A (zh) | 一种图像处理方法、装置及计算机可读存储介质 | |
CN114118127A (zh) | 一种视觉场景标志检测与识别方法及装置 | |
CN112597934A (zh) | 图像目标检测在打包机面积计算中的应用 | |
Liang et al. | Research on Building Extraction from High-Resolution Remote Sensing Image Based on Improved U-Net | |
CN118629029B (zh) | 环境感知方法、装置、电子设备及存储介质 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 23819706 Country of ref document: EP Kind code of ref document: A1 |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2024526384 Country of ref document: JP |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2023819706 Country of ref document: EP |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
ENP | Entry into the national phase |
Ref document number: 2023819706 Country of ref document: EP Effective date: 20250107 |