A Visual Feature-Based End-to-End Method and System for Detecting Backfat Thickness in Pigs

Abstract

This invention proposes an end-to-end method and system for detecting backfat thickness in pigs based on visual features, comprising: acquiring training pig images, wherein each training pig image has a pre-labeled pig region and the backfat thickness of the pig in the image; using the training pig images and their corresponding pig regions, training a multi-scale feature extraction model based on a convolutional neural network to obtain a deep convolutional neural network model for pig instance segmentation, and extracting visual features of the pigs in the training pig images based on the deep convolutional neural network model; using the visual features of the training pig images and their corresponding backfat thickness, training a graph neural network model to obtain a backfat thickness prediction model; inputting the pig image to be detected into the deep convolutional neural network model to obtain the image features of the pig in the pig image to be detected, and inputting them into the backfat thickness prediction model to obtain the backfat thickness of the pig in the pig image to be detected.

Description

Technical Field

[0001] This invention relates to the field of smart farming, and in particular to an end-to-end method and system for detecting backfat thickness in pigs based on visual features. Background Technology

[0002] Backfat thickness helps assess the reproductive performance of sows and the lean meat percentage of fattening pigs. Current technologies for obtaining backfat thickness in pigs often employ either expert-based or ultrasound-based methods. Expert-based methods involve farmers estimating the backfat thickness of live pigs based on their experience. Ultrasound-based methods use ultrasound equipment to scan specific areas of the back of live pigs to detect backfat thickness. Commonly used ultrasound methods include A-mode ultrasound and B-mode ultrasound.

[0003] Testing based on expert experience relies too heavily on the expertise of farmers, resulting in inconsistent results and making it unsuitable for large-scale live pig backfat testing. Ultrasonic testing, when using a pig backfat analyzer, requires fixing the pig and removing hair from specific areas of its back, which can easily cause stress. Furthermore, to avoid human error, testing typically requires a single, dedicated person, which is time-consuming and labor-intensive, significantly limiting testing efficiency. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes an intelligent detection method that reduces reliance on expert prior knowledge and enables rapid and accurate detection of backfat thickness in live pigs. This method solves the problem of large estimation errors and time-consuming and labor-intensive methods for estimating backfat thickness in live pigs under complex farming environments.

[0005] Specifically, this invention proposes an end-to-end method for detecting backfat thickness in pigs based on visual features, comprising:

[0006] Step 1: Obtain multiple training pig images, each of which has a pre-labeled pig region and backfat thickness of the pig in the image.

[0007] Step 2: Using the training pig image and its corresponding pig region, train a multi-scale feature extraction model based on a convolutional neural network to obtain a deep convolutional neural network model for pig instance segmentation, and extract the visual features of the pigs in the training pig image based on the deep convolutional neural network model.

[0008] Step 3: Using the visual features of the training pig images and their corresponding backfat thickness, train the graph neural network model to obtain the backfat thickness prediction model.

[0009] Step 4: Input the image of the pig to be detected into the deep convolutional neural network model to obtain the image features of the pig in the image of the pig to be detected, and input them into the backfat thickness prediction model to obtain the backfat thickness of the pig in the image of the pig to be detected as the detection result.

[0010] The end-to-end pig backfat thickness detection method based on visual features includes each training pig image having a pre-labeled shooting angle; the backfat thickness prediction model includes a feature extraction backbone network, a feature preprocessing module based on a fully connected network, and an adaptive feature semantic mining module.

[0011] Step 3 includes: inputting the pig instance features of each training pig image into the instance feature preprocessing module to map all pig instance features to the same high-dimensional space to form a unified visual feature representation; and using the feature extraction backbone network to process the pig instance features and detect the shooting angle of each training pig image.

[0012] Select the semantic mining module corresponding to the detected shooting angle from the adaptive feature semantic mining module, and predict the backfat thickness based on the unified representation of the visual feature.

[0013] The end-to-end pig backfat thickness detection method based on visual features includes a shooting perspective that includes side view, top view, and rear view, and an adaptive feature semantic mining module that includes a side view feature semantic mining module, a top view feature semantic mining module, and a rear view feature semantic mining module.

[0014] The described end-to-end pig backfat thickness detection method based on visual features involves the pig image to be detected being a multi-view image of the same pig. The multi-view or specified-view images of the pig image to be detected are input into the deep convolutional neural network model to obtain the image features of the multi-view or specified-view images. The image features of the multi-view or specified-view images are then input into the backfat thickness prediction model. The backfat thickness prediction results of each semantic mining module in the adaptive feature semantic mining module of the backfat thickness prediction model are combined to obtain the backfat thickness of the pig in the pig image to be detected.

[0015] This invention also proposes an end-to-end pig backfat thickness detection system based on visual features, comprising:

[0016] The initial module is used to acquire multiple training pig images, and each training pig image has a pre-labeled pig region in the image and the backfat thickness of the pig in the image;

[0017] The feature extraction module is used to train a multi-scale feature extraction model based on a convolutional neural network based on the training pig image and its corresponding pig region, to obtain a deep convolutional neural network model for pig instance segmentation, and to extract the visual features of pigs in the training pig image based on the deep convolutional neural network model.

[0018] The training module is used to train a graph neural network model based on the visual features of the training pig images and their corresponding backfat thickness, so as to obtain a backfat thickness prediction model.

[0019] The prediction module inputs the image of the pig to be detected into the deep convolutional neural network model to obtain the image features of the pig in the image, and then inputs them into the backfat thickness prediction model to obtain the backfat thickness of the pig in the image as the detection result.

[0020] The aforementioned end-to-end pig backfat thickness detection system based on visual features includes each training pig image having a pre-labeled shooting angle; the backfat thickness prediction model includes a feature extraction backbone network, a feature preprocessing module based on a fully connected network, and an adaptive feature semantic mining module.

[0021] The training module is used to: input the pig instance features of each training pig image into the instance feature preprocessing module to map all pig instance features to the same high-dimensional space to form a unified visual feature representation; and use the feature extraction backbone network to process the pig instance features and detect the shooting angle of each training pig image.

[0022] Select the semantic mining module corresponding to the detected shooting angle from the adaptive feature semantic mining module, and predict the backfat thickness based on the unified representation of the visual feature.

[0023] The aforementioned end-to-end pig backfat thickness detection system based on visual features includes a shooting perspective that includes side view, top view, and rear view, and an adaptive feature semantic mining module that includes a side view feature semantic mining module, a top view feature semantic mining module, and a rear view feature semantic mining module.

[0024] The aforementioned end-to-end pig backfat thickness detection system based on visual features includes a pig image to be detected that is a multi-view image of the same pig. The multi-view or specified-view images of the pig image to be detected are input into the deep convolutional neural network model to obtain the image features of the multi-view or specified-view images. The image features of the multi-view or specified-view images are then input into the backfat thickness prediction model. The backfat thickness prediction results of each semantic mining module in the adaptive feature semantic mining module of the backfat thickness prediction model are combined to obtain the backfat thickness of the pig in the pig image to be detected.

[0025] The present invention also proposes a storage medium for storing a program for executing any of the visual feature-based end-to-end pig backfat thickness detection methods.

[0026] The present invention also proposes a client for any of the aforementioned end-to-end pig backfat thickness detection systems based on visual features.

[0027] As can be seen from the above solutions, the advantages of the present invention are:

[0028] This invention provides a software algorithm solution for estimating the backfat thickness of live pigs in the context of smart farming. It eliminates the need for complex hardware installation and configuration. In large-scale pig farming scenarios, based on machine vision algorithms, it effectively addresses the reliance on prior knowledge from farm personnel for backfat thickness estimation in actual farming environments, reducing labor and material costs while minimizing harm to pig health. During the recognition process, an instance segmentation algorithm is applied to accurately extract the contours of live pigs in the image, fully extracting multi-angle visual features valuable for backfat detection, including top-down, side-view, and rear-view angles.

[0029] In training the backfat thickness regression model, the professional knowledge of pig farmers and the detection results of the backfat meter are fully utilized. A backfat thickness regression model based on a convolutional neural network is designed, and regression training is performed on multi-view visual feature data of pigs to achieve backfat thickness detection in pig images from any angle. This enables rapid and accurate backfat thickness estimation in actual farming environments, improving the level of intelligent management in pig farming.

[0030] This invention uses deep learning algorithms to detect the backfat thickness of live pigs, reducing reliance on the prior knowledge of farmers and avoiding estimation errors caused by differences in farmers' physical condition and experience. It utilizes multi-view visual information from live pigs for backfat thickness detection, eliminating the need for hardware to restrict the pigs' activity range and avoiding health problems caused by stress. It also overcomes the limitation of requiring a single person for backfat detection, improving detection speed. Depending on different needs, it can perform single or batch backfat detection on a single live pig or multiple live pigs within unobstructed visual range, ensuring high efficiency in pig backfat detection. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of a multi-view pig instance feature extraction method;

[0032] Figure 2 A schematic diagram of a comprehensive prediction method for backfat thickness in pigs;

[0033] Figure 3 This is a schematic diagram of single-view visual semantic mining and backfat thickness regression prediction. Detailed Implementation

[0034] To address the challenge of detecting backfat thickness in live pigs under smart farming conditions, this invention utilizes multi-view live pig images for instance segmentation and regression prediction of backfat thickness. Regression prediction refers to the process where, although the training data labels are discrete (e.g., backfat thickness 16 mm, 17 mm, etc.), regression prediction outputs a smooth prediction result; that is, given visual features, the final predicted result can be, for example, 17.5 mm, 20.6 mm, etc.

[0035] The present invention mainly includes the following steps: a) Collecting multi-view images of live pigs from the side, top, back, and front, and manually annotating them to construct a multi-view pig image dataset that can be used for live pig instance segmentation and backfat thickness regression analysis. b) Designing visual feature extraction and segmentation techniques for pig image instance segmentation, including but not limited to convolutional neural networks, recurrent neural networks, and graph neural networks, and training an image segmentation model using the constructed multi-view pig images and instance annotation information. c) Designing a visual feature regression model for backfat thickness prediction, and training the visual feature regression model using pig instance features and backfat thickness annotation information. d) During testing, serializing pig instance segmentation and backfat thickness regression techniques to perform instance segmentation and backfat thickness prediction on multi-view live pig images, achieving backfat thickness detection for pig images from any viewpoint.

[0036] To achieve the above-mentioned objectives, this invention designs an end-to-end backfat thickness detection method for pigs based on visual features, comprising the following steps:

[0037] 1. Training of the Pig Instance Segmentation Model. First, a large number of live pig images were collected in a real-world pig farming environment, covering images from different shooting angles such as side view, top view, front view, and rear view. Then, some of the collected live pig images were manually labeled as pig instances to construct a live pig instance segmentation dataset. As an instance segmentation task, a deep convolutional neural network model for live pig instance segmentation was designed and trained on the constructed live pig instance segmentation dataset to extract multi-view visual features of pigs.

[0038] 2. Training of the Pig Backfat Thickness Regression Model. Using the expertise of pig farmers and the detection results from a backfat detector, the backfat thickness of the pig visual features extracted by the pig instance segmentation model was manually labeled, forming a pig backfat thickness regression dataset. A backfat thickness regression model based on multi-view pig visual features was designed and trained on the pig backfat thickness dataset to achieve regression prediction of pig backfat thickness.

[0039] To make the above-mentioned features and effects of the present invention clearer and easier to understand, specific embodiments are provided below, along with detailed descriptions in conjunction with the accompanying drawings. The process of achieving the invention's objective and the corresponding effects are explained in conjunction with the implementation process.

[0040] This invention provides an end-to-end method for detecting backfat thickness in pigs based on visual features. The method flow is shown in the figure, and the specific implementation process includes the following three steps:

[0041] 1. For example Figure 1 As shown, the training of the pig instance segmentation model

[0042] (1) Construction of Live Pig Instance Segmentation Dataset. In a real pig farming environment, live pigs were photographed from multiple perspectives using RGB image acquisition equipment. To obtain multi-angle visual features of the same target, each live pig was photographed from at least three angles, including side view, rear view, and top view. Photography was conducted under various lighting conditions while ensuring the clarity of the RGB images of the live pigs. In this embodiment, approximately 30,000 live pig images were collected, with multiple multi-angle images for each live pig. The original image pixel size was 1920*1080. Each pig image was manually labeled as a pig instance, completing the construction of the live pig instance segmentation dataset. 70% of the images in the dataset were used as the training set, 20% as the validation set, and 10% as the test set.

[0043] The manual annotation steps here include:

[0044] 1) Use a rectangle to mark the area where the target pig is located, that is, obtain the coordinates of the upper left corner and the lower right corner of the rectangle to mark the approximate range of the individual pig;

[0045] 2) Mark multiple points on the pig outline, record the coordinates of each point, and complete the pig instance annotation. Connecting multiple points on the pig outline will mark the pig instance.

[0046] 3) Mark key points on the pig's limbs, including eyes, mouth, neck, elbows, wrists, fingers, and buttocks, and record the coordinate information.

[0047] All annotation information, such as the coordinates of the rectangle, the coordinates of the pig outline points, and the coordinates of the key points of the pig's limbs, is saved to files in txt or json formats.

[0048] (2) Training of the live pig instance segmentation model. Instance segmentation is a popular task in computer vision, essentially involving the accurate identification of targets in visual images and the segmentation of individual targets from the image background at the pixel level. In the live pig instance segmentation task, classic instance segmentation models (such as Mask R-CNN) can be used. End-to-end training of the segmentation model on the pig instance segmentation dataset is crucial, and key steps include:

[0049] Input information: multi-view images of pigs and pixel annotations corresponding to pig instances.

[0050] Model training: Train the model to correctly output the pixel outline of the pig in the image.

[0051] The core of this step is transfer learning and model optimization. Based on the idea of ​​transfer learning, the classic instance segmentation model is first pre-trained on large-scale image datasets such as ImageNet, and then optimized on multi-view pig images to form an instance segmentation technique specifically for pig targets. Model optimization includes, but is not limited to, optimization strategies such as feature communication between different network branches in the instance segmentation model and model pruning.

[0052] (3) Other possible embodiments. In this step, the acquisition of RGB images of live pigs may be carried out under different lighting and farming conditions depending on the actual breeding environment. For the same individual pig, due to the movement characteristics of live pigs, it is necessary to acquire different numbers of pig images from different angles. For the implementation of the pig instance segmentation task in the image, the instance segmentation model used includes, but is not limited to, DeepMask, YOLACT, TensorMask, etc.

[0053] 2. For example Figure 2 As shown, the training of the regression model for backfat thickness in pigs.

[0054] (1) Construction of the regression dataset of live pig backfat thickness. After training the pig instance segmentation model, the backfat thickness of live pigs output by the backfat detector and the shooting angle are used as labeled data. Combined with the visual features of pig instances in each image, the regression dataset of pig backfat thickness is constructed.

[0055] (2) Training of the regression model for backfat thickness in pigs. Using pig instance features as input data and backfat thickness as output, an instance feature preprocessing module based on a fully connected network (FC) was designed to map the input pig instance features to the same high-dimensional space, forming a unified representation of visual features; a feature extraction backbone network was simultaneously used to detect the shooting angle.

[0056] Based on the detected shooting angle, an adaptive "single-view semantic mining module" is selected for backfat thickness prediction, such as side-view feature semantic mining and top-view feature semantic mining. In the "single-view semantic mining module," a multi-scale feature extraction model based on a convolutional neural network (CNN) and a backfat thickness regression model based on a graph neural network (GCN) are designed; these respectively realize the extraction of multi-scale visual features of pigs and the mining of multi-scale visual feature association information; softmax is applied to achieve regression prediction of specific backfat thickness. The structure of the single-view semantic mining module is as follows: Figure 3 As shown, different functions can be achieved based on different pig image observation perspectives, forming... Figure 2 The modules shown include the side-view feature semantic mining module, the top-view feature semantic mining module, and the rear-view feature semantic mining module.

[0057] It is important to note that during the training phase, multi-view visual features of live pig instances are used as input. After training different "single-view semantic mining modules," the testing phase can use either multi-view or single-view visual features as input. If multi-view visual features are used, the outputs of multiple "single-view semantic mining modules" are comprehensively evaluated to achieve a comprehensive prediction of the backfat thickness of the live pig. If single-view visual features are used as input, the output of a single "single-view semantic mining module" is used as the final prediction of backfat thickness.

[0058] (3) Other possible embodiments. In this step, the regression model for backfat thickness of pigs may consist of different numbers of convolutional layers and residual modules, and the feature extraction network used may include, but is not limited to, classical convolutional neural networks, recurrent neural networks and graph neural networks.

[0059] 3. End-to-end prediction of backfat thickness in live pigs in images

[0060] (1) The pig instance segmentation model is used to extract pixel-level visual features of pigs in a given image, and the pig backfat thickness regression model is used to predict the specific value of backfat thickness. In this embodiment, the instance segmentation model and the backfat thickness regression model are sequentially connected, that is, the output of the instance segmentation model is used as the input of the backfat thickness regression model to achieve end-to-end prediction of pig backfat thickness.

[0061] (2) For unlabeled pig images from any viewpoint, input them into the pig instance segmentation model and output the instance features of the pigs; then input them into the backfat thickness regression model to extract multi-scale visual features and mine relational information, and output accurate prediction of backfat thickness.

[0062] (3) Other possible implementations. In the end-to-end backfat thickness detection process, the initial image input may be an RGB image or an RGB+D image containing depth information, as well as the image data acquisition perspective. The output results may be saved to text files such as json, hdf5, and txt, and synchronously mapped onto the visualization interface.

[0063] During the application phase, the initial input to this end-to-end model can be a multi-view image of a single pig sample, or an image of a pig from any viewpoint.

[0064] The following are system embodiments corresponding to the above method embodiments. This embodiment can be implemented in conjunction with the above embodiments. The relevant technical details mentioned in the above embodiments are still valid in this embodiment, and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above embodiments.

[0065] This invention also proposes an end-to-end pig backfat thickness detection system based on visual features, comprising:

[0066] The initial module is used to acquire multiple training pig images, and each training pig image has a pre-labeled pig region in the image and the backfat thickness of the pig in the image;

[0067] The feature extraction module is used to train a multi-scale feature extraction model based on a convolutional neural network based on the training pig image and its corresponding pig region, to obtain a deep convolutional neural network model for pig instance segmentation, and to extract the visual features of pigs in the training pig image based on the deep convolutional neural network model.

[0068] The training module is used to train a graph neural network model based on the visual features of the training pig images and their corresponding backfat thickness, so as to obtain a backfat thickness prediction model.

[0069] The prediction module inputs the image of the pig to be detected into the deep convolutional neural network model to obtain the image features of the pig in the image, and then inputs them into the backfat thickness prediction model to obtain the backfat thickness of the pig in the image as the detection result.

[0070] The aforementioned end-to-end pig backfat thickness detection system based on visual features includes each training pig image having a pre-labeled shooting angle; the backfat thickness prediction model includes a feature extraction backbone network, a feature preprocessing module based on a fully connected network, and an adaptive feature semantic mining module.

[0071] The training module is used to: input the pig instance features of each training pig image into the instance feature preprocessing module to map all pig instance features to the same high-dimensional space to form a unified visual feature representation; and use the feature extraction backbone network to process the pig instance features and detect the shooting angle of each training pig image.

[0072] Select the semantic mining module corresponding to the detected shooting angle from the adaptive feature semantic mining module, and predict the backfat thickness based on the unified representation of the visual feature.

[0073] The aforementioned end-to-end pig backfat thickness detection system based on visual features includes a shooting perspective that includes side view, top view, and rear view, and an adaptive feature semantic mining module that includes a side view feature semantic mining module, a top view feature semantic mining module, and a rear view feature semantic mining module.

[0074] The aforementioned end-to-end pig backfat thickness detection system based on visual features includes a pig image to be detected that is a multi-view image of the same pig. The multi-view or specified-view images of the pig image to be detected are input into the deep convolutional neural network model to obtain the image features of the multi-view or specified-view images. The image features of the multi-view or specified-view images are then input into the backfat thickness prediction model. The backfat thickness prediction results of each semantic mining module in the adaptive feature semantic mining module of the backfat thickness prediction model are combined to obtain the backfat thickness of the pig in the pig image to be detected.

[0075] The present invention also proposes a storage medium for storing a program for executing any of the visual feature-based end-to-end pig backfat thickness detection methods.

[0076] The present invention also proposes a client for any of the aforementioned end-to-end pig backfat thickness detection systems based on visual features.

Claims

1. A method for end-to-end detection of backfat thickness in pigs based on visual features, characterized in that, include: Step 1: Obtain multiple training pig images, each of which has a pre-labeled pig region and backfat thickness of the pig in the image. Step 2: Using the training pig image and its corresponding pig region, train a multi-scale feature extraction model based on a convolutional neural network to obtain a deep convolutional neural network model for pig instance segmentation, and extract the visual features of the pigs in the training pig image based on the deep convolutional neural network model. Step 3: Using the visual features of the training pig images and their corresponding backfat thickness, train the graph neural network model to obtain the backfat thickness prediction model. Step 4: Input the image of the pig to be detected into the deep convolutional neural network model to obtain the image features of the pig in the image of the pig to be detected, and input the image features of the pig in the image of the pig to be detected into the backfat thickness prediction model to obtain the backfat thickness of the pig in the image of the pig to be detected as the detection result; Each training image of a pig also has a pre-labeled shooting angle; the backfat thickness prediction model includes a feature extraction backbone network, a feature preprocessing module based on a fully connected network, and an adaptive feature semantic mining module. Step 3 includes: inputting the pig instance features of each training pig image into the instance feature preprocessing module to map all pig instance features to the same high-dimensional space to form a unified visual feature representation; and using the feature extraction backbone network to process the pig instance features and detect the shooting angle of each training pig image. Select the semantic mining module corresponding to the detected shooting angle from the adaptive feature semantic mining module, and predict the backfat thickness based on the unified representation of the visual feature.

2. The end-to-end pig backfat thickness detection method based on visual features as described in claim 1, characterized in that, The shooting angles include side view, top view, and rear view.

3. The end-to-end pig backfat thickness detection method based on visual features as described in claim 1, characterized in that, The pig image to be detected is a multi-view image of the same pig. The multi-view or specified view image of the pig image to be detected is input into the deep convolutional neural network model to obtain the image features of the multi-view or specified view. The image features of the multi-view or specified view are then input into the backfat thickness prediction model. The backfat thickness prediction results of each semantic mining module in the adaptive feature semantic mining module of the backfat thickness prediction model are combined to obtain the backfat thickness of the pig in the pig image to be detected.

4. An end-to-end pig backfat thickness detection system based on visual features, characterized in that, include: The initial module is used to acquire multiple training pig images, and each training pig image has a pre-labeled pig region in the image and the backfat thickness of the pig in the image; The feature extraction module is used to train a multi-scale feature extraction model based on a convolutional neural network based on the training pig image and its corresponding pig region, to obtain a deep convolutional neural network model for pig instance segmentation, and to extract the visual features of pigs in the training pig image based on the deep convolutional neural network model. The training module is used to train a graph neural network model based on the visual features of the training pig images and their corresponding backfat thickness, so as to obtain a backfat thickness prediction model. The prediction module inputs the image of the pig to be detected into the deep convolutional neural network model to obtain the image features of the pig in the image of the pig to be detected, and inputs the image features of the pig in the image of the pig to be detected into the backfat thickness prediction model to obtain the backfat thickness of the pig in the image of the pig to be detected as the detection result. Each training image of a pig also has a pre-labeled shooting angle; the backfat thickness prediction model includes a feature extraction backbone network, a feature preprocessing module based on a fully connected network, and an adaptive feature semantic mining module. The training module is used to: input the pig instance features of each training pig image into the instance feature preprocessing module to map all pig instance features to the same high-dimensional space to form a unified visual feature representation; and use the feature extraction backbone network to process the pig instance features and detect the shooting angle of each training pig image. Select the semantic mining module corresponding to the detected shooting angle from the adaptive feature semantic mining module, and predict the backfat thickness based on the unified representation of the visual feature.

5. The end-to-end pig backfat thickness detection system based on visual features as described in claim 4, characterized in that, The shooting angles include side view, top view, and rear view.

6. The end-to-end pig backfat thickness detection system based on visual features as described in claim 4, characterized in that, The pig image to be detected is a multi-view image of the same pig. The multi-view or specified view image of the pig image to be detected is input into the deep convolutional neural network model to obtain the image features of the multi-view or specified view. The image features of the multi-view or specified view are then input into the backfat thickness prediction model. The backfat thickness prediction results of each semantic mining module in the adaptive feature semantic mining module of the backfat thickness prediction model are combined to obtain the backfat thickness of the pig in the pig image to be detected.

7. A storage medium for storing a program for executing the end-to-end pig backfat thickness detection method based on visual features as described in any one of claims 1 to 3.

8. A client for the end-to-end pig backfat thickness detection system according to any one of claims 4 to 6.

Images