WO2022209435A1 - コンピュータプログラム、モデル生成方法、推定方法及び推定装置 - Google Patents
コンピュータプログラム、モデル生成方法、推定方法及び推定装置 Download PDFInfo
- Publication number
- WO2022209435A1 WO2022209435A1 PCT/JP2022/007299 JP2022007299W WO2022209435A1 WO 2022209435 A1 WO2022209435 A1 WO 2022209435A1 JP 2022007299 W JP2022007299 W JP 2022007299W WO 2022209435 A1 WO2022209435 A1 WO 2022209435A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- data
- image data
- image
- fish
- aquatic organism
- Prior art date
Links
- 238000000034 method Methods 0.000 title claims abstract description 50
- 238000004590 computer program Methods 0.000 title claims abstract description 39
- 230000011218 segmentation Effects 0.000 claims abstract description 59
- 241000251468 Actinopterygii Species 0.000 claims description 232
- 210000004894 snout Anatomy 0.000 claims description 54
- 238000012937 correction Methods 0.000 claims description 20
- 238000012549 training Methods 0.000 claims description 17
- 238000006073 displacement reaction Methods 0.000 claims description 6
- 238000003384 imaging method Methods 0.000 claims description 6
- 238000012986 modification Methods 0.000 claims 1
- 230000004048 modification Effects 0.000 claims 1
- 238000012545 processing Methods 0.000 abstract description 23
- 238000005259 measurement Methods 0.000 description 51
- 238000010586 diagram Methods 0.000 description 47
- 238000013527 convolutional neural network Methods 0.000 description 10
- 238000010801 machine learning Methods 0.000 description 10
- 238000004364 calculation method Methods 0.000 description 7
- 241001474791 Proboscis Species 0.000 description 4
- 238000013459 approach Methods 0.000 description 3
- 238000006243 chemical reaction Methods 0.000 description 3
- 230000003287 optical effect Effects 0.000 description 3
- 238000012706 support-vector machine Methods 0.000 description 3
- 230000009182 swimming Effects 0.000 description 3
- 208000008589 Obesity Diseases 0.000 description 2
- 210000003323 beak Anatomy 0.000 description 2
- 238000009395 breeding Methods 0.000 description 2
- 230000001488 breeding effect Effects 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 235000020824 obesity Nutrition 0.000 description 2
- 238000009360 aquaculture Methods 0.000 description 1
- 244000144974 aquaculture Species 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 235000013305 food Nutrition 0.000 description 1
- 230000010365 information processing Effects 0.000 description 1
- 230000002093 peripheral effect Effects 0.000 description 1
- 238000003860 storage Methods 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 1
Images
Classifications
-
- A—HUMAN NECESSITIES
- A01—AGRICULTURE; FORESTRY; ANIMAL HUSBANDRY; HUNTING; TRAPPING; FISHING
- A01K—ANIMAL HUSBANDRY; AVICULTURE; APICULTURE; PISCICULTURE; FISHING; REARING OR BREEDING ANIMALS, NOT OTHERWISE PROVIDED FOR; NEW BREEDS OF ANIMALS
- A01K61/00—Culture of aquatic animals
- A01K61/90—Sorting, grading, counting or marking live aquatic animals, e.g. sex determination
- A01K61/95—Sorting, grading, counting or marking live aquatic animals, e.g. sex determination specially adapted for fish
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01B—MEASURING LENGTH, THICKNESS OR SIMILAR LINEAR DIMENSIONS; MEASURING ANGLES; MEASURING AREAS; MEASURING IRREGULARITIES OF SURFACES OR CONTOURS
- G01B11/00—Measuring arrangements characterised by the use of optical techniques
- G01B11/02—Measuring arrangements characterised by the use of optical techniques for measuring length, width or thickness
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/60—Analysis of geometric attributes
- G06T7/62—Analysis of geometric attributes of area, perimeter, diameter or volume
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Definitions
- the present invention relates to computer programs, model generation methods, estimation methods, and estimation devices.
- the size of farmed fish in the cage is important information that determines the amount of feed and the timing of landing.
- multiple analysis services have been provided for automating underwater fish length measurement. These services measure the length by extracting a plurality of characteristic points of the fish body and measuring the distance between the characteristic points (for example, between the end of the mouth and the fork).
- Patent Document 1 discloses an information processing device that detects two characteristic parts (mouth and caudal fin) of a fish from a photographed image of the fish and calculates the size of the fish based on the length between the detected characteristic parts. It is
- the optimal body height measurement for fish does not necessarily have features, so it is not possible to accurately measure body height in terms of measurement accuracy and biological stability. is difficult to estimate.
- the present invention has been made in view of such circumstances, and aims to provide a computer program, a model generation method, an estimation method, and an estimation device capable of estimating the size of aquatic organisms.
- a computer program acquires image data obtained by imaging aquatic organisms into a computer, and outputs position data of a predetermined part of an aquatic organism when the image data is input to a first learning model. to acquire the position data of the predetermined part of the aquatic organism that has been imaged, input the acquired image data into a generation model that generates a segmentation image of the aquatic organism when the image data is input, and A process of generating a segmentation image of the captured aquatic organism and estimating the size of the captured aquatic organism based on the position data of the predetermined part and the segmentation image is executed.
- a computer program provides a second learning model that, when image data is input to a computer, outputs first position data of an aquatic creature and second position data of a predetermined part of the aquatic creature, and stores the obtained image data in a second learning model. to output the first position data and the second position data, to the first learning model is input the image data of the second region including the second position data, and to the generation model , inputting the image data of the first area including the first position data.
- the computer program according to the present invention causes the predetermined part to include the caudal fin and the snout tip, and causes the computer to calculate body height based on a straight line perpendicular to a straight line connecting the caudal fin and the snout tip.
- a computer program inputs image data captured by a stereo camera into the computer into the first learning model, and calculates the three-dimensional positions of the tip of the snout and the tail fork based on the position data of the acquired predetermined part.
- a body height auxiliary line is generated based on the calculated three-dimensional position, and the generated body height auxiliary line is two-dimensionally projected onto the segmentation image to calculate the body height.
- the computer program according to the present invention causes a computer to calculate the body height using the body height auxiliary line located at a predetermined ratio of the distance between the snout end and the fork from the position of the snout end or the fork. , causes the process to be executed.
- the computer program inputs image data of each of a plurality of frames captured by a stereo camera into the computer, and inputs the image data of each of a plurality of frames to the first learning model, and based on the acquired position data of the predetermined part, the tail fork for each frame. calculating the three-dimensional position of the fork, specifying the displacement of the fork for each frame based on the calculated three-dimensional position of the fork, and selecting a frame for estimating the size of the aquatic organism based on the specified displacement , let the process run.
- a computer program displays an estimated position of a caudal fin or a snout tip on a computer based on the position data output by the first learning model, and accepts correction of the estimated position of the displayed caudal fin or snout tip, A process of re-learning the first learning model is executed based on the received correction position and the image data when the correction position is received.
- a computer program displays a segmentation image generated by the generative model on a computer, accepts correction of the displayed segmentation image, and combines the corrected segmentation image and the image data when the correction is accepted. Based on this, a process of re-learning the generative model is executed.
- a computer program according to the present invention causes a computer to execute a process of displaying an image of the aquatic organisms cultivated in a fish tank with fork length or body height attached.
- the computer program according to the present invention divides the accuracy of the estimated size of a plurality of aquatic organisms cultivated in a cage into a plurality of ranks, and displays the number of aquatic organisms for each rank. to run.
- a computer program causes a computer to display a size distribution including at least one of fork length and body height of a plurality of aquatic organisms cultivated in a fish cage.
- a model generation method obtains first training data including image data of captured underwater organisms and position data of predetermined parts of the underwater organisms, and includes the image data and segmentation images of the underwater organisms. Acquiring second training data, and generating a first learning model based on the first training data so as to output position data of a predetermined part of the aquatic organism when image data of the aquatic organism is input. Then, based on the second training data, a generation model is generated so as to generate a segmentation image of the aquatic organism when image data of the aquatic organism is input.
- a model generation method acquires third training data including image data of an aquatic organism, first position data of the aquatic organism, and second position data of a predetermined part of the aquatic organism, A second learning model is generated based on the training data so as to output first position data of the aquatic creature and second position data of a predetermined part of the aquatic creature when image data of the aquatic creature is input. do.
- An estimation method acquires image data of aquatic organisms, and inputs the acquired image data to a first learning model that outputs position data of a predetermined part of an aquatic organism when the image data is input.
- position data of a predetermined portion of the captured aquatic organism is input to a generation model that generates a segmentation image of the aquatic organism when the image data is input;
- a segmentation image of the aquatic organism is generated, and a size of the captured aquatic organism is estimated based on the position data of the predetermined part and the segmentation image.
- the estimating apparatus includes a first acquisition unit that acquires image data of an aquatic organism, and a first learning model that outputs position data of a predetermined part of the aquatic organism when image data is input.
- a second acquisition unit that receives image data acquired by the first acquisition unit and acquires position data of a predetermined portion of the imaged aquatic organism, and a generator that generates a segmentation image of the aquatic organism when the image data is input.
- a generation unit that inputs image data acquired by the first acquisition unit to the model and generates a segmentation image of the captured aquatic organism; and an estimating unit for estimating the size of the aquatic organism.
- FIG. 3 is a schematic diagram showing an example of the configuration of an AI unit
- FIG. 3 is a schematic diagram showing the flow of estimating the size of a fish
- FIG. 10 is a schematic diagram showing an example of the configuration of a second learning model
- FIG. 4 is a schematic diagram showing an example of a configuration of a first learning model
- 1 is a schematic diagram showing an example of a configuration of a generative model
- FIG. 4 is a schematic diagram showing an example of a method of generating each model by machine learning
- FIG. 4 is a schematic diagram showing an example of fish frame pairing between cameras;
- FIG. 3 is a schematic diagram showing an example of the configuration of an AI unit
- FIG. 3 is a schematic diagram showing the flow of estimating the size of a fish
- FIG. 10 is a schematic diagram showing an example of the configuration of a second learning model
- FIG. 4 is a schematic diagram showing an example of a configuration of a first learning model
- 1 is a schematic diagram showing an example of a configuration of
- FIG. 4 is a block diagram showing another example of the configuration of the estimation device according to this embodiment
- FIG. 11 is a schematic diagram showing a first display example of an estimation result
- FIG. 4 is a schematic diagram showing an example of error factors
- FIG. 11 is a schematic diagram showing a second display example of an estimation result
- FIG. 11 is a schematic diagram showing a third display example of an estimation result
- FIG. 11 is a schematic diagram showing a fourth display example of an estimation result
- It is a flowchart which shows an example of the processing procedure of an estimation apparatus. It is a flowchart which shows an example of the processing procedure of an estimation apparatus.
- FIG. 1 is a block diagram showing an example of the configuration of estimation apparatus 100 of this embodiment.
- the estimation device 100 includes an input unit 10 , an AI unit 20 , a pairing unit 30 , a tracking unit 40 , a tail beat removal unit 50 , a measurement unit 60 , an estimation unit 70 as an estimation unit, and an output unit 80 .
- the camera unit 200 is, for example, a waterproof stereo camera installed at a predetermined position underwater in the fish tank.
- the camera unit 200 can capture images of underwater creatures such as fish swimming in a fish tank.
- the camera unit 200 may be a monocular camera. It should be noted that other imaging devices may be used in place of the camera. Below, a fish is mentioned and demonstrated as an example of aquatic life.
- the camera unit 200 outputs captured image data (moving images).
- the captured image data is composed of a plurality of frame-based images (frame images).
- the estimation device 100 may be a personal computer or the like, or a server (cloud) on the Internet. If the estimation device 100 is a personal computer or the like, the estimation device 100 may acquire captured image data directly from the camera section 200 . If the estimating device 100 is a server on the Internet, the estimating device 100 can acquire captured image data captured by the camera unit 200 via a client terminal device (not shown) or the like.
- the input unit 10 acquires captured image data captured by the camera unit 200 .
- the input unit 10 outputs the acquired captured image data to the AI unit 20 .
- the AI unit 20 is a model generated by machine learning, and outputs the data necessary for estimating the size of the fish based on the captured image data.
- Fish size is, for example, fork length and body height.
- the AI unit 20 will be specifically described below.
- the length of the fork is the length of the body, specifically, the length from the tip of the upper jaw (snout) of the fish to the most indented part (fork) of the central part where the caudal fin is bifurcated.
- Body height is the vertical distance from the dorsal edge to the ventral edge of the fish, specifically the distance from the base of the pelvic fin to the dorsal edge.
- FIG. 2 is a schematic diagram showing an example of the configuration of the AI unit 20.
- the AI unit 20 includes a first learning model 21 as a second acquisition unit, a generation model 22 and a second learning model 23 as generation units, and an image clipping unit 24 as a first acquisition unit.
- the estimation unit 70 can estimate the size of the fish based on the position data of the part output by the first learning model 21 and the fish segmentation image data output by the generation model 22 .
- the camera section 200 may be a monocular camera.
- the second learning model 23 outputs position data of a single fish (first position data of aquatic organisms) and position data of parts (second position data of predetermined parts of the aquatic organisms) when captured image data is input. It is generated by performing machine learning to do.
- the position data output by the second learning model 23 are the coordinates (x, y) or the area (x, y, width, height) on the captured image.
- the image cutting unit 24 cuts out the image of the surrounding area including the coordinates or area, and acquires the single fish image data.
- a single fish image includes an image of the entire fish.
- the image clipping unit 24 clips an image of a peripheral area including the coordinates or the area, and acquires part image data.
- the part image includes an image of a predetermined part of the fish.
- the image clipping unit 24 outputs the single fish image data to the generation model 22 and outputs the part image data to the first learning model 21 . Note that the image clipping unit 24 may be provided outside the AI unit 20 .
- the first learning model 21 is generated by performing machine learning so as to output the position data of the part when the part image data output by the image clipping unit 24 is input.
- the generation model 22 is generated by performing machine learning so that when the single fish image data output by the image clipping unit 24 is input, fish segmentation image data is output.
- Fig. 3 is a schematic diagram showing the flow of fish size estimation.
- the second learning model 23 cooperates with the image clipping unit 24 to generate single fish image data (image data of the first region including the first position data) when captured image data is input, and part image data (image data of the second region including the second position data).
- the single fish image is an image of a rectangular area surrounding the entire single fish.
- Single fish image data is data that includes rectangular area data and a class (for example, fish) within the rectangular area.
- the site images are a snout image including the snout end P1 as a predetermined site, and a caudal fin image including a caudal fin (or fork) P2 as a predetermined site.
- the part image data is data including rectangular area data and classes within the rectangular area (for example, snout and caudal fin).
- the first learning model 21 outputs the position data of the part when the part image data is input.
- the positional data of the parts are the positional data of the snout end P1 and the positional data of the caudal fin P2, and are coordinates on the image. If the image coordinate system is XY, the position data can be represented by coordinates (x, y) or area (x, y, width, height). In this case, the coordinate along the optical axis of the camera can be represented by Z.
- the generative model 22 outputs segmentation image data when single fish image data is input.
- the segmentation image is obtained by classifying each pixel (pixel) of the single fish image into classes. In the example of FIG. Pixels classified into are represented by "patterned".
- a segmentation image can be used to determine the boundary between a fish body and a non-fish body (outline of a fish body).
- the estimation unit 70 can estimate the size of the fish based on the position data of the parts and the segmentation image data. As shown in FIG. 3, the length of the straight line L connecting the snout end P1 and the caudal fin P2 can be estimated as the fork length. Also, the length between the intersections of the straight line H perpendicular to the straight line L and the outline of the fish can be estimated as the body height.
- the position of the intersection of the straight line H and the straight line L can be a point of a predetermined percentage on the straight line L from the position of the proboscis P1 (for example, 40% of the length of the fork).
- the position data of the predetermined parts (nose end and caudal fin) and the segmentation image are combined, and body height measurement can be performed based on the body height straight line H orthogonal to the straight line representing the fork length and the segmentation image. Therefore, it is possible to stably and highly accurately estimate the size of a fish even if there is no characteristic part that is optimal for body height measurement.
- body height measurement can be performed with even higher accuracy by performing the processing after the pairing unit 30, which will be described later.
- FIG. 4 is a schematic diagram showing an example of the configuration of the second learning model 23.
- the second learning model 23 can be configured by, for example, Faster R-CNN (Convolutional Neural Network).
- the second learning model 23 comprises a CNN layer 231 , a region proposal network (RPN) 232 , a ROI (Region-of-Interest) POOL 233 and an identification network 234 .
- the CNN layer 231 generates an image feature amount (feature map) from the input captured image data and outputs it to the region proposal network (RPN) 232 and ROI POOL 233 .
- a region proposal network (RPN) 232 calculates candidate regions from the input image feature values and outputs them to the ROI POOL 233 .
- a region proposal network (RPN) 232 can detect where an object appears in the captured image, that is, the area in which the object appears and its rectangular shape.
- the ROI pool 233 connects the image feature quantity output from the CNN layer 231 and the candidate regions output from the region proposal network (RPN) 232, and outputs fixed-length ROI region features to the identification network 234.
- RPN region proposal network
- the identification network 234 recalculates the correct area/class from the input ROI area features, classifies the area and the class of the area, and outputs it.
- the classes are solitary fish, snout and caudal fin.
- the second learning model 23 is not limited to Faster R-CNN, for example, R-CNN, Mask R-CNN, YOLO (You Only Look Once), SSD (Single Shot Multibox Detector), etc. good.
- FIG. 5 is a schematic diagram showing an example of the configuration of the first learning model 21.
- the first learning model 21 can be composed of, for example, RetinaNet.
- the first learning model 21 comprises a Feature Pyramid Network 211 , a classifier 212 and a region regressor 213 .
- the part image data output by the second learning model 23 is input to the first learning model 21 .
- the region image data includes caudal fin image data and snout tip image data.
- Feature Pyramid Network 211 computes feature hierarchies that compose feature maps of various scales in the bottom-up direction, and in the top-down direction, it uploads feature maps from higher layers that are spatially coarse but semantically strong. Sample to generate high-resolution features. Each feature is associated with features computed in a bottom-up direction.
- the Feature Pyramid Network 211 can extract feature quantities that have both high-order and low-order characteristics, and can achieve well-balanced accuracy in terms of both meaning and location.
- the region regressor 213 outputs the position of the snout end and the position of the caudal fin.
- the positions of the snout and caudal fin are represented by coordinate values on the image.
- the class classifier 212 identifies the class (proboscis and caudal fin) of the positions (coordinate values) output by the region regressor 213 .
- the position of the snout tip on the image of the snout tip can be obtained with high accuracy
- the position of the caudal fin on the image of the caudal fin can be obtained with high accuracy.
- the first learning model 21 is not limited to RetinaNet, and may be, for example, SegNet, Mask R-CNN, SVM (Support Vector Machine), or the like.
- FIG. 6 is a schematic diagram showing an example of the configuration of the generative model 22.
- the generative model 22 can be composed of U-Net, for example.
- the generative model 22 includes encoders 221-225 and decoders 226-229.
- the generative model 22 repeats convolution processing with encoders 221 to 225 on the input single fish image data. Decoders 226 to 229 repeat upsampling (deconvolution) of the image convoluted by encoder 225 .
- decoding the convolved image processing is performed to add the feature maps generated by the encoders 224 to 221 to the image to be deconvolved.
- the generative model 22 outputs a segmentation image of the fish on the input single fish image.
- a segmentation image can be used to determine the boundary between a fish body and a non-fish body (outline of a fish body).
- the generative model 22 is not limited to U-Net, and may be, for example, GAN (Generative Adversarial Network), SegNet, or the like.
- FIG. 7 is a schematic diagram showing an example of a method for generating each model by machine learning.
- FIG. 7A shows a method of generating the second learning model 23.
- the second learning model 23 performs machine learning so as to output position data of a single fish (first position data) and position data of parts (second position data) when captured image data is input as input data for learning. You can go and generate it.
- position data of a single fish and position data of parts created by annotating based on captured image data, which is input data for learning are used as teacher data.
- a large number of learning data and teacher data are prepared, and the second learning model 23 is machine-learned.
- the internal parameters of the second learning model 23 can be updated to generate the second learning model 23 so that the output data approaches the teacher data.
- FIG. 7B shows a method of generating the first learning model 21.
- the first learning model 21 can be generated by performing machine learning so that when part image data is input as input data for learning, position data of the part is output. In this case, positional data of parts created by annotating parts based on part image data, which is input data for learning, is used as teacher data. A large number of learning data and teacher data are prepared, and the first learning model 21 is machine-learned. The internal parameters of the first learning model 21 can be updated to generate the first learning model 21 so that the output data approaches the teacher data.
- FIG. 7C shows how the generative model 22 is generated.
- the generation model 22 can be generated by performing machine learning so as to output segmentation image data when single fish image data is input as input data for learning.
- the teacher data uses segmentation image data created by performing annotation based on the single fish image data, which is the input data for learning.
- a large number of learning data and teacher data are prepared, and the generative model 22 is machine-learned.
- the generative model 22 can be generated by updating the internal parameters of the generative model 22 so that the output data approaches the teacher data.
- Each of the models mentioned above can also be retrained. For example, an operation of displaying the estimated positions of the caudal fin and the snout tip on the screen based on the position data output by the first learning model 21 as shown in FIG. 3 and correcting the displayed estimated positions of the caudal fin and the snout tip. accept. Correction of the estimated position is performed, for example, by operating a mouse or the like to move the pointer on the screen to the estimated position, and performing a predetermined operation (touch, click, drag and drop, etc.) to change the estimated position on the screen. can be moved to the correct position.
- a predetermined operation touch, click, drag and drop, etc.
- the received corrected position and the image data when the corrected position is received are prepared as training data, and the first learning model 21 is regenerated based on the received corrected position and the image data when the corrected position is received. can learn. This improves the accuracy of the position data output by the first learning model 21 .
- the segmentation image generated by the generation model 22 is displayed on the screen, and an operation for correcting the displayed segmentation image is accepted. Correction of the segmentation image is performed, for example, by starting an application such as paint software, displaying the segmentation image on the screen, operating the mouse or the like to move the pointer on the screen to a desired position, and performing a predetermined operation (touch, Click, drag and drop, etc.) to convert non-fish pixels into fish pixels, or convert fish pixels into non-fish pixels.
- the corrected segmentation image and the image data when the correction is accepted are prepared as training data, and the generative model 22 is re-learned based on the corrected segmentation image and the image data when the correction is accepted. be able to. This improves the accuracy of the segmentation image generated by the generative model 22 .
- the position data of the single fish, the position data of the snout tip, and the position data of the caudal fin output by the second learning model 23 are displayed on the screen, and an operation for correcting the displayed position data is accepted.
- clipped images single fish image, snout end image, caudal fin image
- the position of the single fish output by the second learning model 23 is displayed on the screen, and an operation for correcting the position of the displayed single fish is accepted.
- Correction of the position of the single fish may be performed, for example, by operating the mouse or the like to move the position of the single fish on the screen and correcting the position.
- the position of the snout tip output by the second learning model 23 is displayed on the screen, and an operation for correcting the displayed position of the snout tip is accepted.
- Correction of the position of the proboscis may be performed, for example, by operating a mouse or the like to move the position of the proboscis on the screen and correcting the position.
- the position of the caudal fin output by the second learning model 23 is displayed on the screen, and an operation for correcting the position of the displayed caudal fin is accepted. Correction of the position of the caudal fin may be performed, for example, by operating a mouse or the like to move the position of the caudal fin on the screen and correcting the position. Further, position data corrected for re-learning the first learning model 21 can also be used as the position data of the snout tip and caudal fin.
- the corrected position of the single fish and the image data when the correction is accepted are prepared as training data, and a second learning model 23 is generated based on the corrected position of the single fish and the image data when the correction is accepted. can be re-learned.
- the corrected position of the rostrum and the image data at the time of receiving the correction are prepared as training data, and second learning is performed based on the corrected position of the rostrum and the image data at the time of receiving the correction.
- Model 23 can be retrained.
- the corrected caudal fin position and the image data when the correction is accepted are prepared as training data, and based on the corrected caudal fin position and the image data when the correction is accepted, the second The learning model 23 can be retrained. As a result, the accuracy of the position data of the single fish and the position data of the part output by the second learning model 23 is improved.
- the pairing unit 30 performs pairing processing between cameras and pairing processing between fish and parts. First, pairing processing between cameras will be described. Pairing processing between cameras includes fish frame pairing and body part frame pairing.
- Fig. 8 is a schematic diagram showing an example of fish frame pairing between cameras.
- the left identification image is an image showing a single fish image estimated based on the image captured by the left camera of the stereo camera
- the right identification image is captured by the right camera of the stereo camera.
- This is an image showing a single fish image (also referred to as a fish frame) estimated based on the picked-up image.
- three fish frames are illustrated in the example of FIG. 8, the number of fish frames is not limited to three in an actual image.
- the stereo cameras in the left-right direction horizontal direction
- they may be arranged in the vertical direction, or may be arranged in an oblique direction (for example, an angle greater than 0 degrees and less than 90 degrees with respect to the horizontal direction). You may be arranged in the left-right direction (horizontal direction), they may be arranged in the vertical direction, or may be arranged in an oblique direction (for example, an angle greater than 0 degrees and less than 90 degrees with respect to the horizontal direction). You may be arranged
- the pairing unit 30 performs round-robin pairing for all fish frames in the left and right identification images.
- pairing is performed between the fish frame G1 in the left identification image and all the fish frames G4, G5, and G6 in the right identification image.
- the fish frame G6 is selected as the closest pair to the fish frame G1.
- a corresponding pair is selected from the fish frames in the right side identification image.
- body part frame pairing can also be performed in the same manner as fish frame pairing.
- the pairing unit 30 performs round-robin pairing for all body part frames in the left and right identification images. Pairing is performed between the body part frame of the left identification image and all the body part frames in the right identification image. When pairing, select the closest pair within a tolerance.
- FIG. 9 is a diagram showing an example of a pair list of a fish frame and a body part frame.
- FIG. 9A shows a fish frame pair list
- FIG. 9B shows a body part frame pair list.
- the fish frame G1 of the left identification image and the fish frame G6 of the right identification image are paired.
- the fish frame G2 of the left identification image and the fish frame G4 of the right identification image are paired
- the fish frame G3 of the left identification image and the fish frame G5 of the right identification image are paired.
- the body part frame g1 of the left identification image and the body part frame g12 of the right identification image are paired, the body part frame g2 of the left identification image and the body part frame g10 of the right identification image are paired, and the left identification image and the body part frame g8 of the right identification image are paired.
- FIG. 10 is a schematic diagram showing an example of pairing fish and parts. Assume that the fish frames G1 to G3 and the body part frames g1 to g6 are paired by the pairing between the cameras described above. The fish frames G1 to G3 and the body part frames g1 to g6 each have a pair of fish frames and body part frames. FIG. 10 illustrates one side (for example, the left side identification image) of the paired fish frame and body part frame.
- the pairing unit 30 performs round-robin pairing for all fish frames and body part frames in the identification image.
- pairing can be performed using only the left identification image, only the right identification image, or both left and right identification images.
- pairing is performed between the fish frame G1 and all body part frames g1 to g6 in the identification image.
- select the closest pair within a tolerance select the closest pair within a tolerance. Similar processing is performed for the other fish frames G2 and G3.
- Fig. 11 is a diagram showing an example of a fish list.
- a fish frame G1 and body part frames g3 and g4 are paired.
- the fish frame G2 and body part frames g1 and g2 form a pair
- the fish frame G3 and body part frames g5 and g6 form a pair. That is, a fish with an ID of 1 is paired with the fish frame G1 and body part frames g3 and g4, a fish with an ID of 2 is paired with a fish frame G2 and body part frames g1 and g2, and a fish with an ID of 3 is , a fish frame G3 and body part frames g5 and g6 are paired.
- FIG. 12 is a block diagram showing an example of the configuration of the tracking unit 40.
- the tracking unit 40 tracks the fish swimming in the fish tank between frames. Double counting of fish can be prevented by tracking.
- the tracking unit 40 includes a 2D tracking unit 41 , a 3D conversion unit 42 and a 3D tracking unit 43 .
- the 2D tracking unit 41 uses the fish frame pair list, body part frame pair list, and fish list generated by the pairing unit 30 to track each fish.
- the 2D tracking unit 41 performs tracking based on the 2D image of the fish.
- the 3D conversion unit 42 converts the single fish image obtained by imaging with the left camera and the image with the right camera. 2D images of the fish are generated by using the single fish image obtained by the method and the ranging principle of the stereo camera. Convert to D image. Alternatively, the 3D positions of the snout tip and fork may be measured based on the principle of triangulation.
- the 3D tracking unit 43 uses the fish frame pair list, body part frame pair list, and fish list generated by the pairing unit 30 to track each fish.
- the 3D tracking unit 43 can track based on the position of the snout tip on the 3D image of the fish. This can reduce the effect of the fish's tail beat when the fish swims in the water.
- FIG. 13 is a block diagram showing an example of the configuration of the tail beat removal section 50.
- the tail beat removal unit 50 includes a fork 3D measurement unit 51 , a plane projection unit 52 , an approximate straight line calculation unit 53 and a frame removal unit 54 .
- the fork 3D measurement unit 51 measures the 3D position (X, Y, Z) from the (X, Y) coordinates on the image of the fork (caudal fin) paired by the pairing unit 30 .
- the horizontal direction of the image can be the X axis
- the vertical direction can be the Y axis
- the optical axis direction of the camera can be the Z axis.
- the plane projection unit 52 projects the 3D position of the fork onto the XZ plane. As a result, the position of the fork can be the position seen from the back side of the fish.
- the plane projection unit 52 records the position of the fork seen from the back side of the fish for each frame. This allows the position of the fork to be plotted on the XZ plane.
- the approximate straight line calculation unit 53 calculates an approximate straight line of the fork position (fork trajectory) plotted on the XZ plane.
- the frame removing unit 54 removes the frame corresponding to the fork when the deviation of the fork from the approximate straight line exceeds the allowable range. Instead of calculating the approximate straight line after projecting the 3D position of the fork onto a plane, an approximate line in the 3D space may be calculated from the 3D position of the fork. In this case, if the deviation of the 3D position of the fork from the calculated 3D approximation line exceeds the allowable range, the frame corresponding to the fork should be removed.
- FIG. 14 is a schematic diagram showing an example of an approximate straight line of the fork trajectory.
- the horizontal axis indicates the position X of the fork
- the vertical axis indicates the position Z of the fork.
- O is the measured position of the fork
- symbols t1 to t13 indicate the time points of the frames
- the dashed straight line indicates the approximate straight line of the fork trajectory.
- the position of the fork changes from time t1 to t13.
- the positions of the forks at times t3 and t7 to t9 deviate from the approximate straight line.
- the frames at times t3 and t7 to t9 are removed, and the remaining frames at times t1 to t2, t4 to t6, and t10 to t13 are used to measure body height, which will be described later.
- the position of the fork deviates from the approximation line because, for example, the fish changes course and the fork fluctuates greatly. If the fork deviation exceeds the allowable range, the frame corresponding to the fork is removed.
- Frame removal is not limited to the method using the approximate straight line, and other machine learning such as SVM, neural network, clustering, etc. may be used. Also, if the amount of temporal change in the position of the fork exceeds a predetermined threshold, the frame may be removed.
- the tail beat removal unit 50 outputs the frames other than the frames from which the tail beat has been removed among the imaging frames to the measurement unit 60 as tail beat removed frames (frames to be measured).
- the estimating apparatus 100 inputs the image data of each of a plurality of frames captured by the stereo camera to the first learning model 21, and based on the acquired position data of the predetermined part, the three-dimensional position of the fork for each frame. is calculated, the calculated three-dimensional position is projected onto a predetermined two-dimensional plane, the displacement of the fork is specified for each frame, and a frame for estimating the size of the fish is selected based on the specified displacement.
- FIG. 15 is a block diagram showing an example of the configuration of the measuring section 60.
- the measurement unit 60 includes an optimum frame selection unit 61 , a snout/fork 3D measurement unit 62 , a body height auxiliary line generation unit 63 , a body height auxiliary line projection unit 64 , and a body height measurement unit 65 .
- the optimum frame selection unit 61 selects the frame at the time corresponding to the fork closest to the approximate straight line of the fork trajectory. Instead of selecting the optimum frame, it may be randomly selected from the remaining frames after frame removal, or a few frames before or after the removed frame may be selected. good.
- the snout/fork 3D measurement unit 62 measures the 3D position of the snout and the 3D position of the fork in the selected frame. For the 3D measurement, the 3D position of the snout end tracked by the 3D tracking unit 43 and the 3D position of the fork measured by the fork 3D measurement unit 51 may be used.
- the body height auxiliary line generation unit 63 divides the 3D straight line connecting the 3D position of the 3D measured snout tip and the 3D position of the tail fork into 10, for example, and divides the 3D straight line into nine body height auxiliary lines perpendicular to the 3D straight line. to generate
- the body height auxiliary line projection unit 64 projects the body height auxiliary line generated by the body height auxiliary line generation unit 63 onto a 2D plane (for example, the XY plane).
- the body height measurement unit 65 measures the body height based on the position of intersection between the body height auxiliary line projected onto the 2D plane and the contour of the fish (for example, the contour on the segmentation image). Note that if the body height cannot be measured due to a defect in the contour of the fish, the optimum frame selection unit 61 selects the point of time corresponding to the fork closest to the approximate straight line of the fork trajectory. A frame can be selected, and each process of the snout/fork 3D measuring unit 62, the body height auxiliary line generating unit 63, the body height auxiliary line projecting unit 64, and the body height measuring unit 65 can be repeated.
- Defects in the contour of the fish can be determined, for example, by comparing the distance between the intersections of the auxiliary body height line and the contour of the fish with a predetermined threshold. Alternatively, determination can be made by comparing the amount of change in the distance between the intersections of the adjacent auxiliary body height lines and the outline of the fish with a predetermined threshold value. Missing portions may be interpolated.
- FIG. 16 is a schematic diagram showing an example of body height measurement.
- the 3D straight line L connecting the 3D position of the 3D-measured snout end and the 3D position of the tail fork is divided into, for example, 10, and nine body height auxiliary lines orthogonal to the 3D straight line L Generate A1-A9.
- the body height auxiliary lines A1 to A9 are projected onto a 2D plane (for example, the XY plane), and the body height auxiliary lines A1 to A9 projected onto the 2D plane and the contour of the fish (for example, on the segmentation image)
- the height is measured based on the position of intersection with the contour of the body.
- the body height can be measured as the distance H between two positions where the body height auxiliary line A4 at a distance corresponding to 40% of the fork length L from the snout end intersects the outline of the fish body.
- the number of body height auxiliary lines is an example, and is not limited to nine.
- the ratio of 40% is an example and is not limited to 40%.
- FIG. 17 is a schematic diagram showing an example of automatic measurement of fish size.
- FIG. 17 shows IDs of fish appearing in images of respective frame numbers. For example, frames 1 to 10 show a fish with a fish ID of 0001, and in frame 7 indicated by O, the size of fish 0001 is automatically measured.
- Frame 7 measured by automatic measurement is, for example, the frame with the smallest tail beat among frames 1-10.
- the fish with ID 0002 is shown in frames 3 to 12, the size of fish 0002 is automatically measured in frame 9 indicated by O, and the fish with ID 0003 is shown in frames 5 to 9.
- the size of fish 0003 is not automatically measured.
- Frames 7 to 14 show a fish with a fish ID of 0004.
- Frames 8 to 16 show a fish with a fish ID of 0005.
- frame 15 indicated by O the size of fish 0005 is automatically measured.
- the tracking unit 40 recognizes the same individual between frames.
- the measurement unit 60 outputs the measurement result to the estimation unit 70.
- the estimation unit 70 can collect the automatically measured fork length and body height of the fish in the cage and estimate the size of the fish in the cage.
- the output unit 80 can convert the result estimated by the estimation unit 70 into displayable data and output the data to an external terminal device, display device, or the like.
- FIG. 18 is a schematic diagram showing an example of the configuration of the estimation result DB 90.
- the estimation result DB 90 may be provided inside the estimation device 100 or may be provided in an external data server or the like as long as it is accessible from the estimation device 100 .
- the estimation result DB 90 can store estimation results from the estimation device 100 .
- the estimation result DB 90 records, for example, fish IDs, fork lengths, body heights, ranks, and images in association with fish IDs.
- position information (coordinate values) of the snout tip and position information (coordinate values) of the caudal fin (fork) may be recorded in association with the fish ID.
- a fish ID is, for example, an identifier that identifies a fish in a fish tank.
- the rank specifies the measurement accuracy of the size of the fish. Details of the rank will be described later.
- the image is an image of a fish, and may be superimposed on the image of the fish to display a straight line indicating the fork length and a straight line indicating the body height of the fish. Also, the frame number, fork length value, body height value, and rank to which the fish belongs may be displayed on the image of the fish.
- FIG. 19 is a block diagram showing another example of the configuration of estimation device 100 of the present embodiment.
- the estimation device 100 can use, for example, a personal computer or the like.
- the estimation device 100 can be configured with a CPU 101, a ROM 102, a RAM 103, a GPU 104, a video memory 105, a recording medium reading unit 106, and the like.
- a computer program (computer program product) recorded on a recording medium 1 for example, an optically readable disk storage medium such as a CD-ROM
- a recording medium reading unit 106 for example, an optical disk drive
- the computer program includes processing procedures described in FIGS. 25 and 26 described later. It may be stored in a hard disk (not shown) and stored in the RAM 103 when the computer program is executed.
- the CPU 101 By causing the CPU 101 to execute a computer program (computer program product) stored in the RAM 103, the input unit 10, the AI unit 20, the pairing unit 30, the tracking unit 40, the tail beat removal unit 50, the measurement unit 60, the estimation unit 70 , and each process in the output unit 80 can be executed.
- the video memory 105 can temporarily store data and processing results for various image processing.
- the computer program (computer program product) can be downloaded from another computer, network device, or the like via a network such as the Internet instead of being read by the recording medium reading unit 106 .
- the estimation result displayed on the display screen of the terminal device connected to the estimation device 100 via the communication network will be described.
- the following display processing is performed according to the request from the terminal device. This can be done by having the estimation device 100 access the estimation result DB 90, and if necessary, calculate numerical values for display and output them to the terminal device. Further, when the terminal device has already downloaded the data of the estimation result DB, the data downloaded by the CPU of the terminal device can be used to display on the display unit.
- FIG. 20 is a schematic diagram showing a first display example of estimation results.
- the estimation result screen 300 displays columns such as rank, number of measurements, average fish weight, average fork length, average body height, average degree of obesity, average photographing distance, and average angle. Also, a box 301 for receiving selection of download of detailed data is displayed for each rank, and detailed data can be downloaded from the estimation device 100 by operating a “save” icon 302 .
- the data to be downloaded includes moving images captured by the camera unit 200 .
- rank A has an error of less than 5%
- rank B has an error of 5% or more and less than 10%
- rank C has an error of 10%.
- rank F can be defined as having an error of 20% or more. Note that the number of rank divisions and the definition of the rank are not limited to these.
- FIG. 21 is a schematic diagram showing an example of error factors.
- FIG. 21A shows the distribution of site location errors. As the error in the position of the snout and the error in the position of the tail fork (caudal fin) increase, the errors in the fork length and body height increase. The error of the position of the snout and the fork is calculated based on the reprojection error when three-dimensional measurement is performed using the position data of the snout and the caudal fin (fork) output by the first learning model 21. .
- FIG. 21B shows the inclination (angle) error of the fish body. The greater the inclination of the fish body, the greater the body height error. The inclination error can be reflected in the body height error by converting the angle deviation into a simple measurement position deviation. The ranking may employ the worst of the site position errors or tilt errors.
- the number of measurements indicates the number of fish included in each rank.
- Average fork length and average body height represent average values within each rank.
- the average fish weight and average obesity degree can be calculated from the average fork length and average body height by a predetermined formula.
- the weight of the breeding fish can be converted from the fork length and body height estimated by the estimation device 100 without using an actual weight scale.
- the body height can be estimated stably and with high accuracy even if there is no characteristic part that is optimal for body height measurement, so the weight of the breeding fish can also be estimated with high accuracy.
- the estimating device 100 classifies the accuracy of the estimated size of a plurality of fishes (aquatic organisms) cultured in a fish tank into a plurality of ranks, and displays the fish for each classed rank. As a result, the size of the fish in the cage and the reliability of the data can be determined, and information for determining the feeding amount and landing timing can be provided.
- FIG. 22 is a schematic diagram showing a second display example of the estimation result.
- a distribution selection list 311 and a distribution display area 312 are displayed on the estimation result screen 310 .
- fork length distribution is selected from items such as fork length, body height, and fish weight, and the distribution display area 312 displays the fork length distribution of fish in the cage.
- rank may be displayed as a breakdown of the distribution of fork lengths. Similar to the fork length, the distribution of body height and fish weight can be displayed.
- the estimating device 100 can display the distribution of at least one of the fork length and body height of a plurality of fish cultured in a fish cage. This allows us to see how the sizes of the fish are distributed in the pen, which can provide information for determining the amount of food to be fed and the timing of landing.
- FIG. 23 is a schematic diagram showing a third display example of the estimation result.
- a rank selection list 321 and a fish measurement data display area 322 are displayed on the estimation result screen 320 .
- rank A is selected from ranks A, B, C, and F.
- the measurement data display area 322 the measurement data of each fish belonging to rank A is displayed.
- the measurement data includes, but is not limited to, fish weight, fork length, body height, and images.
- FIG. 24 is a schematic diagram showing a fourth display example of the estimation result.
- An image of the selected fish is displayed on the estimation result screen 330 .
- a straight line indicating the fork length and a straight line indicating the body height of the fish may be displayed by being superimposed on the image of the fish.
- the frame number, fork length value, body height value, fish weight value, and rank to which the fish belongs may be displayed on the image of the fish.
- the estimation device 100 can display the image of the fish with the fork length and body height attached. This makes it possible to check the measurement data of each fish (individual) belonging to the rank.
- the estimation device 100 acquires captured image data (S11), inputs the acquired captured image data to the second learning model 23, and acquires single fish image data and part image data from the image clipping unit 24 (S12 ).
- the region image data includes snout image data and caudal fin image data.
- the estimating device 100 inputs the acquired part image data to the first learning model 21 to acquire position data of the part (S13).
- the snout tip image data is input to the first learning model 21, the position information (coordinates or region values) of the snout tip output by the first learning model 21 is acquired, and the caudal fin image data is sent to the first learning model 21. to acquire the caudal fin position information (coordinates or region values) output by the first learning model 21 .
- the estimation device 100 inputs the acquired single fish image data to the generation model 22 to acquire segmentation image data (S14), and estimates the fork length measurement point based on the acquired position data of the part and the segmentation image data. (Generate) (S15).
- measurement points for fork length positional information (coordinates or region values) of each of the snout and caudal fin
- the size of the fish may be estimated by obtaining the distance Z of the fish or a predetermined portion of the fish from the camera using the processes of steps S16 and S17 described later.
- the estimating device 100 performs pairing between cameras (S16), pairs fish and parts, and generates a fish list (S17).
- the estimation device 100 performs fish tracking processing (S18) and performs tail beat removal processing (S19).
- the estimating device 100 selects the optimum frame for body height calculation (S20).
- the estimating device 100 determines whether the body height can be calculated (S21), and if the body height cannot be calculated (NO in S21), the process of step S20 is performed. If the body height can be calculated (YES in S21), the estimating device 100 calculates the body height (S22).
- the body height is calculated by the method shown in FIG. 16B. "Calculation" of body height is synonymous with “measurement”, but it also means to measure indirectly.
- the estimating device 100 determines the presence or absence of other fish (fish that have not been measured) (S23).
- the estimation device 100 repeats the processes from step S11 onwards. If there is no other fish (NO in S23), the estimation device 100 collects fish calculation data (S24), and classifies the calculation data into ranks according to the error (S25). The estimating device 100 calculates a statistical value (e.g., average value) of the calculated data for each rank (S26), creates a distribution for each estimation item (e.g., fork length, body height, etc.) (S27), and obtains an estimation result. is output (S28), and the process ends.
- a statistical value e.g., average value
- the estimation device 100 can be configured with a CPU, GPU, ROM, RAM, recording medium reading unit, and the like.
- a computer program recorded on a recording medium can be read by a recording medium reading unit and stored in a RAM.
- the processes performed by the estimation device 100 can be executed by causing the CPU and GPU to execute computer programs stored in the RAM.
- the computer program can be downloaded via a network such as the Internet instead of being read by the recording medium reading unit.
- the body height measurement is performed by combining the segmentation image and the body height straight line, so even if there is no optimum part for body height measurement, the size of the fish can be stably and highly accurately estimated.
- the fins of fish expand and retract during swimming, which causes a decrease in measurement accuracy.
- segmentation is performed on the fish except for the fins, so the body height can be measured with high accuracy without being affected by the fins.
- extracting feature points from high-resolution images with high accuracy increases the computational load for image processing.
- the size of the fish can be estimated with high accuracy by combining the segmentation image and the body height line, and the load can be reduced. can.
- estimation device 10 input unit 20 AI unit 21 first learning model 211 Feature Pyramid Network 212 class classifier 213 region regressor 22 generative model 221, 222, 223, 224, 225 encoder 226, 227, 228, 229 decoder 23 second learning model 24 image clipping unit 231 CNN layer 232 region proposal network 233 ROI POOL 234 identification network 30 pairing unit 40 tracking unit 41 2D tracking unit 42 3D conversion unit 43 3D tracking unit 50 tail beat removal unit 51 fork 3D measurement unit 52 plane projection unit 53 approximate straight line calculation unit 54 frame removal unit 60 measurement unit 61 Optimal frame selection unit 62 snout/fork 3D measurement unit 63 body height auxiliary line generation unit 64 body height auxiliary line projection unit 65 body height measurement unit 70 estimation unit 80 output unit 90 estimation result DB 101 CPUs 102 ROMs 103 RAM 104 GPUs 105 video memory 106 recording medium reading unit 200 camera unit
Landscapes
- Life Sciences & Earth Sciences (AREA)
- Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Zoology (AREA)
- Environmental Sciences (AREA)
- General Physics & Mathematics (AREA)
- Marine Sciences & Fisheries (AREA)
- Animal Husbandry (AREA)
- Biodiversity & Conservation Biology (AREA)
- Geometry (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
- Length Measuring Devices By Optical Means (AREA)
Abstract
Description
うに第2学習モデルを生成する。
る。
座標はZで表すことができる。
とができる。生成モデル22は、U-Netに限定されるものではなく、例えば、GAN(Generative Adversarial Network)、SegNet等でもよい。
D画像に変換する。または、吻端及び尾叉の3D位置を3点測量の原理に基づき計測してもよい。魚の2D画像の各画素までの距離Zは、Z=(B×F)/Dで算出することができる。ここで、Bはカメラ間距離であり、Fは焦点距離であり、Dは視差である。
ぞれの画像データを第1学習モデル21に入力して、取得した所定部位の位置データに基づいてフレーム毎の尾叉の3次元位置を算出し、算出した3次元位置を所定の2次元平面に投影してフレーム毎に尾叉の変位を特定し、特定した変位に基づいて、魚のサイズを推定するためのフレームを選択する。
で、例えば、最もテールビートが小さいフレームである。
推定装置100が推定結果DB90にアクセスし、必要な場合には、表示するための数値を演算し、端末装置へ出力することにより行うことができる。また、端末装置が、推定結果DBのデータを既にダウンロードしている場合には、端末装置のCPUがダウンロードしたデータを用いて表示部に表示することにより行うことができる。
10 入力部
20 AI部
21 第1学習モデル
211 Feature Pyramid Network
212 クラス分類器
213 領域回帰器
22 生成モデル
221、222、223、224、225 エンコーダ
226、227、228、229 デコーダ
23 第2学習モデル
24 画像切出部
231 CNN層
232 領域提案ネットワーク
233 ROI POOL
234 識別ネットワーク
30 ペアリング部
40 追尾部
41 2D追尾部
42 3D変換部
43 3D追尾部
50 テールビート除去部
51 尾叉3D計測部
52 平面投影部
53 近似直線算出部
54 フレーム除去部
60 計測部
61 最適フレーム選定部
62 吻端・尾叉3D計測部
63 体高補助線生成部
64 体高補助線投影部
65 体高計測部
70 推定部
80 出力部
90 推定結果DB
101 CPU
102 ROM
103 RAM
104 GPU
105 ビデオメモリ
106 記録媒体読取部
200 カメラ部
Claims (15)
- コンピュータに、
水中生物を撮像した画像データを取得し、
画像データを入力した場合に水中生物の所定部位の位置データを出力する第1学習モデルに、取得した画像データを入力して、前記撮像された水中生物の所定部位の位置データを取得し、
画像データを入力した場合に水中生物のセグメンテーション画像を生成する生成モデルに、取得した画像データを入力して、前記撮像された水中生物のセグメンテーション画像を生成し、
前記所定部位の位置データ及び前記セグメンテーション画像に基づいて前記撮像された水中生物のサイズを推定する、
処理を実行させるコンピュータプログラム。 - コンピュータに、
画像データを入力した場合に水中生物の第1位置データ及び前記水中生物の所定部位の第2位置データを出力する第2学習モデルに、取得した画像データを入力して、前記第1位置データ及び前記第2位置データを出力し、
前記第1学習モデルには、前記第2位置データを含む第2領域の画像データを入力し、
前記生成モデルには、前記第1位置データを含む第1領域の画像データを入力する、
処理を実行させる請求項1に記載のコンピュータプログラム。 - 前記所定部位は、尾鰭及び吻端を含み、
コンピュータに、
前記尾鰭と前記吻端とを結ぶ直線と直交する直線に基づいて体高を算出する、
処理を実行させる請求項1又は請求項2に記載のコンピュータプログラム。 - コンピュータに、
ステレオカメラによって撮像された画像データを前記第1学習モデルに入力して、取得した所定部位の位置データに基づいて吻端及び尾叉の3次元位置を算出し、
算出した3次元位置に基づいて体高補助線を生成し、
生成した体高補助線を前記セグメンテーション画像上に2次元投影して体高を算出する、
処理を実行させる請求項1から請求項3のいずれか一項に記載のコンピュータプログラム。 - コンピュータに、
前記吻端又は前記尾叉の位置から、前記吻端と前記尾叉の距離の所定割合の位置にある前記体高補助線を用いて前記体高を算出する、
処理を実行させる請求項4に記載のコンピュータプログラム。 - コンピュータに、
ステレオカメラによって撮像された複数のフレームそれぞれの画像データを前記第1学習モデルに入力して、取得した所定部位の位置データに基づいてフレーム毎の尾叉の3次元位置を算出し、
算出した尾叉の3次元位置に基づいてフレーム毎に尾叉の変位を特定し、
特定した変位に基づいて、前記水中生物のサイズを推定するためのフレームを選択する、
処理を実行させる請求項1から請求項5のいずれか一項に記載のコンピュータプログラム。 - コンピュータに、
前記第1学習モデルが出力する位置データに基づいて、尾鰭又は吻端の推定位置を表示し、
表示した尾鰭又は吻端の推定位置の修正を受け付け、
受け付けた修正位置と、前記修正位置を受け付けた際の画像データとに基づいて、前記第1学習モデルを再学習する、
処理を実行させる請求項1から請求項6のいずれか一項に記載のコンピュータプログラム。 - コンピュータに、
前記生成モデルが生成したセグメンテーション画像を表示し、
表示したセグメンテーション画像の修正を受け付け、
修正されたセグメンテーション画像と、前記修正を受け付けた際の画像データとに基づいて、前記生成モデルを再学習する、
処理を実行させる請求項1から請求項7のいずれか一項に記載のコンピュータプログラム。 - コンピュータに、
生簀内で養殖されている前記水中生物の画像に尾叉長又は体高を付して表示する、
処理を実行させる請求項1から請求項8のいずれか一項に記載のコンピュータプログラム。 - コンピュータに、
生簀内で養殖されている複数の水中生物の推定したサイズの精度を複数のランクに区分し、区分したランク毎の水中生物の数を表示する、
処理を実行させる請求項1から請求項9のいずれか一項に記載のコンピュータプログラム。 - コンピュータに、
生簀内で養殖されている複数の水中生物の尾叉長及び体高の少なくとも一方を含むサイズの分布を表示する、
処理を実行させる請求項1から請求項10のいずれか一項に記載のコンピュータプログラム。 - 水中生物を撮像した画像データ、及び前記水中生物の所定部位の位置データを含む第1訓練データを取得し、
前記画像データ、及び前記水中生物のセグメンテーション画像を含む第2訓練データを取得し、
前記第1訓練データに基づいて、水中生物を撮像した画像データを入力した場合に、前記水中生物の所定部位の位置データを出力するように第1学習モデルを生成し、
前記第2訓練データに基づいて、水中生物を撮像した画像データを入力した場合に、前記水中生物のセグメンテーション画像を生成するように生成モデルを生成する、
モデル生成方法。 - 水中生物を撮像した画像データ、前記水中生物の第1位置データ及び前記水中生物の所定部位の第2位置データを含む第3訓練データを取得し、
前記第3訓練データに基づいて、水中生物を撮像した画像データを入力した場合に、前記水中生物の第1位置データ及び前記水中生物の所定部位の第2位置データを出力するように第2学習モデルを生成する、
請求項12に記載のモデル生成方法。 - 水中生物を撮像した画像データを取得し、
画像データを入力した場合に水中生物の所定部位の位置データを出力する第1学習モデルに、取得した画像データを入力して、前記撮像された水中生物の所定部位の位置データを取得し、
画像データを入力した場合に水中生物のセグメンテーション画像を生成する生成モデルに、取得した画像データを入力して、前記撮像された水中生物のセグメンテーション画像を生成し、
前記所定部位の位置データ及び前記セグメンテーション画像に基づいて前記撮像された水中生物のサイズを推定する、
推定方法。 - 水中生物を撮像した画像データを取得する第1取得部と、
画像データを入力した場合に水中生物の所定部位の位置データを出力する第1学習モデルに、前記第1取得部で取得した画像データを入力して、前記撮像された水中生物の所定部位の位置データを取得する第2取得部と、
画像データを入力した場合に水中生物のセグメンテーション画像を生成する生成モデルに、前記第1取得部で取得した画像データを入力して、前記撮像された水中生物のセグメンテーション画像を生成する生成部と、
前記所定部位の位置データ及び前記セグメンテーション画像に基づいて前記撮像された水中生物のサイズを推定する推定部と
を備える推定装置。
Priority Applications (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
EP22779660.4A EP4317910A1 (en) | 2021-03-31 | 2022-02-22 | Computer program, model generation method, estimation method and estimation device |
CN202280024241.2A CN117178161A (zh) | 2021-03-31 | 2022-02-22 | 计算机程序、模型生成方法、推断方法以及推断装置 |
JP2023510653A JPWO2022209435A1 (ja) | 2021-03-31 | 2022-02-22 |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2021-060624 | 2021-03-31 | ||
JP2021060624 | 2021-03-31 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2022209435A1 true WO2022209435A1 (ja) | 2022-10-06 |
Family
ID=83455891
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2022/007299 WO2022209435A1 (ja) | 2021-03-31 | 2022-02-22 | コンピュータプログラム、モデル生成方法、推定方法及び推定装置 |
Country Status (5)
Country | Link |
---|---|
EP (1) | EP4317910A1 (ja) |
JP (1) | JPWO2022209435A1 (ja) |
CN (1) | CN117178161A (ja) |
CL (1) | CL2023002753A1 (ja) |
WO (1) | WO2022209435A1 (ja) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20230177668A1 (en) * | 2021-12-02 | 2023-06-08 | X Development Llc | Underwater camera biomass prediction aggregation |
JP7413583B1 (ja) | 2023-03-31 | 2024-01-15 | 株式会社電通 | 魚の品質判定システム |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH01195364A (ja) * | 1988-01-29 | 1989-08-07 | Hitachi Ltd | 魚体生態監視装置 |
WO2018061925A1 (ja) | 2016-09-30 | 2018-04-05 | 日本電気株式会社 | 情報処理装置、長さ測定システム、長さ測定方法およびプログラム記憶媒体 |
JP2019003554A (ja) * | 2017-06-19 | 2019-01-10 | コニカミノルタ株式会社 | 画像認識装置、画像認識方法、および画像認識装置用プログラム |
WO2019188506A1 (ja) * | 2018-03-26 | 2019-10-03 | 日本電気株式会社 | 情報処理装置、物体計測システム、物体計測方法およびプログラム記憶媒体 |
JP6842100B1 (ja) * | 2019-08-28 | 2021-03-17 | ウミトロン ピーティーイー エルティーディー | 水棲動物検出装置、情報処理装置、端末装置、水棲動物検出システム、水棲動物検出方法、及び水棲動物検出プログラム |
-
2022
- 2022-02-22 CN CN202280024241.2A patent/CN117178161A/zh active Pending
- 2022-02-22 WO PCT/JP2022/007299 patent/WO2022209435A1/ja active Application Filing
- 2022-02-22 EP EP22779660.4A patent/EP4317910A1/en active Pending
- 2022-02-22 JP JP2023510653A patent/JPWO2022209435A1/ja active Pending
-
2023
- 2023-09-14 CL CL2023002753A patent/CL2023002753A1/es unknown
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH01195364A (ja) * | 1988-01-29 | 1989-08-07 | Hitachi Ltd | 魚体生態監視装置 |
WO2018061925A1 (ja) | 2016-09-30 | 2018-04-05 | 日本電気株式会社 | 情報処理装置、長さ測定システム、長さ測定方法およびプログラム記憶媒体 |
JP2019003554A (ja) * | 2017-06-19 | 2019-01-10 | コニカミノルタ株式会社 | 画像認識装置、画像認識方法、および画像認識装置用プログラム |
WO2019188506A1 (ja) * | 2018-03-26 | 2019-10-03 | 日本電気株式会社 | 情報処理装置、物体計測システム、物体計測方法およびプログラム記憶媒体 |
JP6842100B1 (ja) * | 2019-08-28 | 2021-03-17 | ウミトロン ピーティーイー エルティーディー | 水棲動物検出装置、情報処理装置、端末装置、水棲動物検出システム、水棲動物検出方法、及び水棲動物検出プログラム |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20230177668A1 (en) * | 2021-12-02 | 2023-06-08 | X Development Llc | Underwater camera biomass prediction aggregation |
US11842473B2 (en) * | 2021-12-02 | 2023-12-12 | X Development Llc | Underwater camera biomass prediction aggregation |
JP7413583B1 (ja) | 2023-03-31 | 2024-01-15 | 株式会社電通 | 魚の品質判定システム |
Also Published As
Publication number | Publication date |
---|---|
JPWO2022209435A1 (ja) | 2022-10-06 |
EP4317910A1 (en) | 2024-02-07 |
CL2023002753A1 (es) | 2024-03-08 |
CN117178161A (zh) | 2023-12-05 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US11913771B2 (en) | Information processing device, object measuring system, object measuring method, and program storing medium | |
US10643347B2 (en) | Device for measuring position and orientation of imaging apparatus and method therefor | |
US20210232924A1 (en) | Method for training smpl parameter prediction model, computer device, and storage medium | |
CN110568447B (zh) | 视觉定位的方法、装置及计算机可读介质 | |
WO2022209435A1 (ja) | コンピュータプログラム、モデル生成方法、推定方法及び推定装置 | |
KR101519940B1 (ko) | 3-차원 물체 모델링, 피팅 및 트래킹 | |
CN111028350B (zh) | 一种利用双目立体相机构建栅格地图的方法 | |
CN114424250A (zh) | 结构建模 | |
JP2016099982A (ja) | 行動認識装置、行動学習装置、方法、及びプログラム | |
KR20100038168A (ko) | 구도 해석 방법, 구도 해석 기능을 구비한 화상 장치, 구도 해석 프로그램 및 컴퓨터 판독 가능한 기록 매체 | |
US20160239632A1 (en) | Anatomically Specific Movie Driven Medical Image Review | |
CN112927279A (zh) | 一种图像深度信息生成方法、设备及存储介质 | |
JP2019114103A (ja) | 物体認識処理装置、物体認識処理方法及びプログラム | |
JP2010282615A (ja) | 3dワーピング手法と固有対象物運動(pom)の検出の組み合わせに基づく対象物の動き検出システム | |
CN111178193A (zh) | 一种车道线的检测方法、检测装置及计算机可读存储介质 | |
JP2020013560A (ja) | 情報処理装置、情報処理方法、及びプログラム | |
CN108648145B (zh) | 图像拼接方法及装置 | |
CN111709269B (zh) | 一种深度图像中基于二维关节信息的人手分割方法和装置 | |
CN117726747A (zh) | 补全弱纹理场景的三维重建方法、装置、存储介质和设备 | |
JP2006113832A (ja) | ステレオ画像処理装置およびプログラム | |
CN115546027A (zh) | 图像缝合线确定方法、装置以及存储介质 | |
JP2010187130A (ja) | カメラ校正装置、カメラ校正方法、カメラ校正プログラムおよびそのプログラムを記録した記録媒体。 | |
CN115035070A (zh) | 一种确定cbct图像中种植体位置的方法及装置 | |
JP2002008014A (ja) | 3次元形状抽出方法及び装置並びに記録媒体 | |
JP2013104660A (ja) | 単眼カメラ画像によるステレオ視に基づく運動量推定方法、及び当該方法を使用する運動量推定装置 |
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: 22779660 Country of ref document: EP Kind code of ref document: A1 |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2023510653 Country of ref document: JP |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2022779660 Country of ref document: EP |
|
ENP | Entry into the national phase |
Ref document number: 2022779660 Country of ref document: EP Effective date: 20231031 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |