A chicken cecal content moisture prediction method based on vision and multi-task learning
By employing a vision-based and multi-task learning approach, the cecal contents region is automatically located and cropped. By combining segmentation and regression tasks, the problems of light interference and spatial heterogeneity are solved, enabling robust prediction of chicken cecal contents humidity and multi-site output, thus improving the accuracy and reliability of the measurement.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA AGRI UNIV SANYA RES INST
- Filing Date
- 2026-05-07
- Publication Date
- 2026-06-02
AI Technical Summary
Existing technologies for measuring the humidity of chicken cecal contents suffer from problems such as light interference, loss of spatial heterogeneity, mean regression error, and lack of uncertainty quantification, leading to measurement bias and data instability.
A vision-based and multi-task learning approach is adopted. The target detection module automatically locates and crops the cecal contents region. Combined with segmentation and regression tasks, mask-guided feature aggregation and quantile regression are used to output multiple humidity quantiles, providing interval references.
It achieves improved robustness and deployability of humidity prediction while reducing background interference, accurately characterizes internal regional differences, and provides multi-quantile results that facilitate quality control, as well as screening of abnormal samples and subsequent statistical analysis.
Smart Images
Figure CN122135018A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of computer vision, deep learning and intelligent measurement of animal phenotypes, and specifically relates to a method for predicting the humidity of chicken cecal contents based on vision and multi-task learning. Background Technology
[0002] In poultry breeding and gut health assessment, cecal contents moisture is a crucial phenotype reflecting digestion and absorption, intestinal barrier function, and microbial fermentation status. Traditional moisture measurement typically relies on offline experimental procedures such as weighing and drying, which are time-consuming and difficult to perform simultaneously on a large scale. Image-based moisture estimation methods offer an alternative for rapid phenotyping, but existing techniques generally suffer from the following shortcomings: Background and illumination interference: Cecal contents co-occur with non-target information such as mesentery, surrounding tissues, gloves, dissecting instruments, and local occlusion. This information may interfere with the model's extraction of humidity-related features such as the color, texture, and surface condition of the contents, thus causing prediction bias.
[0003] Loss of spatial heterogeneity: Single global regression often ignores the differences between the upper and lower or center-edge regions within the sample, making it difficult to reflect the true humidity distribution.
[0004] Mean regression and extreme sample error: Point estimation regression often shrinks towards the mean at both ends of the data distribution (too dry / too wet), reducing the usability of the extreme segments.
[0005] Lack of uncertainty quantification: Outputting only a single value cannot measure the degree of confidence, which is not conducive to handling error propagation in quality control and downstream statistics (such as microbial / genetic analysis). Summary of the Invention
[0006] This invention aims to provide a method for predicting the humidity of chicken cecal contents based on vision and multi-task learning, without requiring manual fine-tuning and subjective evaluation: It enables automatic localization and cropping of cecal contents regions to reduce background interference; it jointly learns segmentation and regression tasks in the same network, explicitly introduces mask-guided feature aggregation to improve the effectiveness of region representation, outputs multiple humidity quantiles through quantile regression to provide interval references for humidity prediction results, facilitates abnormal sample screening and result quality control, and outputs standardized structured results to facilitate large-scale collection, quality control, summary statistics and traceability.
[0007] To achieve the above objectives, this invention presents a method for predicting the humidity of chicken cecal contents based on vision and multi-task learning. The method is characterized by the following steps: S1: Collect images of chicken cecal contents, preprocess the images to obtain image data, and construct an image dataset; S2: Construct a humidity prediction model for chicken cecal contents to obtain the humidity estimation results of chicken cecal contents; the humidity prediction model for chicken cecal contents includes a target detection module and a multi-task joint prediction module; S21: Input the image data into the target detection module to obtain the content region, and perform local cropping based on the content region to obtain the ROI cropped image; S22: Input the ROI cropped image into the multi-task joint prediction module to obtain the humidity multi-quantile result; the multi-task joint prediction module includes a feature encoding layer, a segmentation layer, a mask-guided weighted pooling layer, and a quantile regression layer; S23: Process and summarize the humidity multi-site results to obtain the humidity estimation results of the chicken cecal contents; S3: Input the image data into the chicken cecal contents humidity prediction model to obtain the chicken cecal contents humidity estimation result.
[0008] Preferably, in step S1, an image of the contents of the chicken cecum is acquired, and the image of the chicken cecum contents is preprocessed to obtain image data. The specific process is as follows: Images of chicken cecal contents are collected, and the images are subjected to size standardization, color correction, noise suppression, and quality screening. The sample number, image number, and metadata of the collection batch are recorded to obtain image data.
[0009] Preferably, the specific process of obtaining the ROI cropped image in step S21 is as follows: Image data is input into the target detection module, which uses a built-in public detection network to perform forward inference and global traversal to identify the contents of the chicken cecum. This yields a batch of candidate target detection boxes and the predicted confidence level for each candidate box. Based on a set fixed confidence threshold, candidate target detection boxes with confidence levels below the threshold are removed, and false detection interference areas are filtered to obtain target detection boxes. Non-maximum suppression algorithm is used to eliminate overlapping and redundant target detection boxes on the same target, thus obtaining the content region. The pixel coordinates of each content region are read, and local cropping is performed on the image data based on the pixel coordinates. Single cropped images are generated one by one, and metadata such as target detection box coordinates, detection confidence level, and original image number are saved simultaneously. Finally, the ROI cropped image is obtained.
[0010] Preferably, the specific process for obtaining the humidity multi-site results in step S22 is as follows: The ROI-cropped image is input into a feature encoding layer for deep feature extraction to obtain a feature map. The feature map is then input into a segmentation layer to obtain a pixel-level segmentation mask. The pixel-level segmentation mask and the feature map are input into a weighted pooling layer to aggregate and obtain a global representation vector. This global representation vector is then input into a quantile regression layer for mapping and regression to obtain the humidity quantile results. The mask-guided weighted pooling layer uses the following formula:
[0011] in, Represents the global representation vector. Represents a pixel-level segmentation mask. Representing feature maps, This indicates element-wise multiplication.
[0012] Preferably, the specific process for obtaining the moisture estimation result of the chicken cecal contents in step S23 is as follows: The humidity quantile results were scaled or calibrated to restore the moisture content percentage corresponding to the drying and weighing method. The results of multiple humidity quantiles were summarized using the same image or the same sample to obtain the humidity estimation results of the chicken cecal contents.
[0013] The advantages of this invention compared to existing technologies are as follows: By reducing background and non-target area interference through detection and cropping, the robustness and deployability of predictions are improved.
[0014] By using segmentation mask-guided MAP aggregation, we avoid the indiscriminate incorporation of background information by global average pooling, thereby improving the expression quality of effective region features.
[0015] By using quantile regression to output multiple humidity quantiles, interval references are provided for the prediction results, which facilitates quality control, screening of abnormal samples, and subsequent statistical analysis.
[0016] By using robust aggregation of multiple ROIs and optional spatial partitioning, phenotypic characterization of regional differences within a sample can be achieved. Attached Figure Description
[0017] Figure 1 This is a flowchart of a method for predicting the humidity of chicken cecal contents based on vision and multi-task learning according to the present invention. Detailed Implementation
[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments and accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] like Figure 1 As shown, the present invention proposes a method for predicting the humidity of chicken cecal contents based on vision and multi-task learning, which includes the following steps: S1: Collect images of chicken cecal contents, preprocess the images to obtain image data, and construct an image dataset; This includes collecting images of the contents of the chicken cecum and recording metadata associated with the samples; Images of chicken cecal contents are collected, and the images are subjected to size standardization, color correction, noise suppression, and quality screening. The sample number, image number, and metadata of the collection batch are recorded to obtain image data.
[0020] S2: Construct a humidity prediction model for chicken cecal contents to obtain the humidity estimation results of chicken cecal contents; the humidity prediction model for chicken cecal contents includes a target detection module and a multi-task joint prediction module; S21: Input the image data into the target detection module to obtain the content region, and perform local cropping based on the content region to obtain the ROI cropped image; The process involves inputting image data into the target detection module, which uses a built-in public detection network to perform forward inference and global traversal to identify the contents of a chicken cecum. This yields a batch of candidate target detection boxes and the predicted confidence level for each box. Based on a set fixed confidence threshold, candidate target detection boxes with confidence levels below the threshold are removed, and false positive interference areas are filtered to obtain the target detection boxes. A non-maximum suppression algorithm is used to eliminate overlapping and redundant target detection boxes on the same target, thus obtaining the content region. The pixel coordinates of each content region are read, and local cropping is performed on the image data based on the pixel coordinates. Single cropped images are generated one by one, and metadata such as the target detection box coordinates, detection confidence level, and original image number are saved simultaneously. Finally, the ROI cropped image is obtained.
[0021] S22: Input the ROI cropped image into the multi-task joint prediction module to obtain the humidity multi-quantile result; the multi-task joint prediction module includes a feature encoding layer, a segmentation layer, a mask-guided weighted pooling layer, and a quantile regression layer; In the multi-task joint prediction stage, the ROI cropped image is input into the feature encoding layer to extract the image feature map; the image feature map is input into the segmentation layer to obtain a pixel-level segmentation mask for representing the effective region of the cecal contents; the pixel-level segmentation mask and the image feature map are input into the mask to guide the weighted pooling layer to perform weighted aggregation of the image features of the effective region of the contents to obtain a global representation vector; the global representation vector is input into the quantile regression layer to output the humidity multi-quantile result.
[0022] Crop the ROI image and scale it to a preset input size (e.g., 224×224 or other sizes).
[0023] The input feature encoding layer shares a visual feature encoder (either a CNN or a Transformer encoder) to obtain feature maps. .
[0024] The input segmentation layer outputs a pixel-level mask (soft mask or binary mask) to define the effective area of the cecal contents.
[0025] Masked Guided Weighted Pooling (MAP) layers aggregate feature maps into global representation vectors. The mask-guided weighted pooling layer has the following formula:
[0026] in, Represents the global representation vector. Represents a pixel-level segmentation mask. Representing feature maps, Indicates element-wise multiplication; The global representation vector is input into the quantile regression layer, which maps and regresses the fused global representation vector, simultaneously completing multi-branch prediction and outputting humidity multi-quantile results (q). 10 q 50 q 90 (or other quantile combinations).
[0027] S23: Process and summarize the humidity multi-site results to obtain the humidity estimation results of the chicken cecal contents; Perform inverse normalization on the output humidity multi-site results to restore them to moisture content percentage, in % or mass fraction.
[0028] For multiple humidity quantile results of images cropped from the same ROI, grouping them by image number, sample number, or ROI number, and using robust statistics such as median / truncated mean for image-level aggregation, we obtain image-level q. 10 q 50 q 90 and n_crops.
[0029] Output ROI-level and image-level structured files (CSV / JSON), i.e., the moisture estimation results of chicken cecal contents, and record intermediate results for traceability; The training process of the chicken cecal contents humidity prediction model is as follows: Labeling: For object detection tasks, label the bounding boxes of cecal contents; for segmentation tasks, label the pixel-level masks of the contents. COCO format or equivalent labeling formats can be used.
[0030] Training objectives: The segmentation branch can use loss methods such as Dice / BCE; the quantile regression branch uses Pinball Loss to optimize multiple humidity quantiles separately; multi-task weighted joint optimization can be performed.
[0031] Extreme sample strategy: Divide the samples into low / medium / high intervals according to humidity quantiles, and perform balanced sampling in mini-batch to alleviate mean regression in extreme segments.
[0032] Cross-validation or ensemble: K-fold cross-validation is used to train multiple models for ensemble inference to improve generalization and stability.
[0033] S3: Input the image data into the chicken cecal contents humidity prediction model to obtain the chicken cecal contents humidity estimation result.
[0034] Based on the chicken cecal contents humidity prediction model, the humidity of the chicken cecal contents image to be tested is predicted to obtain the chicken cecal contents humidity estimation result.
[0035] Although preferred embodiments of this application have been described, those skilled in the art, once they learn the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application. Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if these modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include these modifications and variations.
Claims
1. A method for predicting the humidity of chicken cecal contents based on vision and multi-task learning, characterized in that, include: S1: Collect images of chicken cecal contents, preprocess the images to obtain image data, and construct an image dataset; S2: Construct a humidity prediction model for chicken cecal contents to obtain the humidity estimation results of chicken cecal contents; the humidity prediction model for chicken cecal contents includes a target detection module and a multi-task joint prediction module; S21: Input the image data into the target detection module to obtain the content region, and perform local cropping based on the content region to obtain the ROI cropped image; S22: Input the ROI cropped image into the multi-task joint prediction module to obtain the humidity multi-quantile result; the multi-task joint prediction module includes a feature encoding layer, a segmentation layer, a mask-guided weighted pooling layer, and a quantile regression layer; S23: Process and summarize the humidity multi-site results to obtain the humidity estimation results of the chicken cecal contents; S3: Input the image data into the chicken cecal contents humidity prediction model to obtain the chicken cecal contents humidity estimation result.
2. The method for predicting the humidity of chicken cecal contents based on vision and multi-task learning according to claim 1, characterized in that, In step S1, an image of the contents of the chicken cecum is acquired, and the image is preprocessed to obtain image data. The specific process is as follows: Images of chicken cecal contents are collected, and the images are subjected to size standardization, color correction, noise suppression, and quality screening. The sample number, image number, and metadata of the collection batch are recorded to obtain image data.
3. The method for predicting the humidity of chicken cecal contents based on vision and multi-task learning according to claim 1, characterized in that, In step S21, the specific process of obtaining the ROI cropped image is as follows: Image data is input into the target detection module, which uses a built-in public detection network to perform forward inference and global traversal to identify the contents of the chicken cecum. This yields a batch of candidate target detection boxes and the predicted confidence level for each candidate box. Based on a set fixed confidence threshold, candidate target detection boxes with confidence levels below the threshold are removed, and false detection interference areas are filtered to obtain target detection boxes. Non-maximum suppression algorithm is used to eliminate overlapping and redundant target detection boxes on the same target, thus obtaining the content region. The pixel coordinates of each content region are read, and local cropping is performed on the image data based on the pixel coordinates. Single cropped images are generated one by one, and metadata such as target detection box coordinates, detection confidence level, and original image number are saved simultaneously. Finally, the ROI cropped image is obtained.
4. The method for predicting the humidity of chicken cecal contents based on vision and multi-task learning according to claim 1, characterized in that, In step S22, the specific process for obtaining the humidity multi-site results is as follows: The ROI-cropped image is input into a feature encoding layer for deep feature extraction to obtain a feature map. The feature map is then input into a segmentation layer to obtain a pixel-level segmentation mask. The pixel-level segmentation mask and the feature map are input into a weighted pooling layer to aggregate and obtain a global representation vector. This global representation vector is then input into a quantile regression layer for mapping and regression to obtain the humidity quantile results. The mask-guided weighted pooling layer uses the following formula: , in, Represents the global representation vector. Represents a pixel-level segmentation mask. Representing feature maps, This indicates element-wise multiplication.
5. The method for predicting the humidity of chicken cecal contents based on vision and multi-task learning according to claim 1, characterized in that, In step S23, the specific process for obtaining the estimated humidity result of the chicken cecal contents is as follows: The humidity quantile results were scaled or calibrated to restore the moisture content percentage corresponding to the drying and weighing method. The results of multiple humidity quantiles were summarized using the same image or the same sample to obtain the humidity estimation results of the chicken cecal contents.
6. A computer-readable storage medium, characterized in that, Used to store computer-readable programs or instructions, which, when executed by a processor, can implement the steps in the method for predicting the humidity of chicken cecal contents based on vision and multi-task learning as described in any one of claims 1 to 5.