Search

CN-121999325-A - Intelligent inspection method for farm based on machine vision

CN121999325ACN 121999325 ACN121999325 ACN 121999325ACN-121999325-A

Abstract

The invention relates to an intelligent inspection method for a farm based on machine vision, which comprises the steps of collecting machine vision data of a plurality of inspection areas in the farm, wherein the machine vision data comprise thermal imaging images and inspection videos, detecting target individuals by utilizing the inspection videos, distributing data IDs to the target individuals, mapping the data IDs to the thermal imaging images, and carrying out anomaly detection on the thermal imaging images with digital ID labels to obtain abnormal body temperature conditions of the target individuals. The invention not only can improve the inspection efficiency and reduce manual intervention, but also can obviously improve the level of animal health management and provide powerful support for the development of modern breeding industry.

Inventors

  • ZHAN GUANGJIAN
  • Qiu Chunxu

Assignees

  • 农翼(苏州)智慧科技有限公司

Dates

Publication Date
20260508
Application Date
20260211

Claims (8)

  1. 1. A machine vision-based intelligent inspection method for a farm is characterized by comprising the following steps: collecting machine vision data of a plurality of inspection areas in a farm, wherein the machine vision data comprises thermal imaging images and inspection videos; Detecting a target individual by using the inspection video, distributing a data ID for the target individual, and mapping the data ID to the thermal imaging image; And carrying out anomaly detection on the thermal imaging image with the digital ID tag to obtain the abnormal body temperature of the target individual.
  2. 2. The machine vision-based intelligent inspection method of a farm of claim 1, wherein determining machine vision data for an inspection area in the farm comprises: a wheeled robot is adopted to carry out inspection in an inspection area of the farm, and the wheeled robot is provided with a thermal imager and a video recorder; the thermal imager is used for acquiring a thermal imaging image of a target in the inspection area; The video recorder is used for collecting the inspection video of the inspection area.
  3. 3. The machine vision-based intelligent inspection method of a farm of claim 1, wherein assigning a data ID to the target individual comprises: dividing the inspection video into a plurality of sections, extracting multi-frame images from each section, stacking the multi-frame images according to a time sequence, and extracting key frames as effective images; inputting the effective image into YOLOv models to obtain a boundary box of the target individual; extracting different features from the target individual detected in the bounding box, wherein the different features comprise appearance features and deep learning features; And assigning a data ID to the target individual by using the different characteristics.
  4. 4. A machine vision based intelligent inspection method of a farm according to claim 3, wherein assigning data IDs to the target individuals using the different features comprises: And carrying out preliminary matching based on the appearance characteristics, screening out target candidate results, carrying out fine matching on the target candidate results through the deep learning characteristics, associating the target with a corresponding known target track if the matching is successful, using the digital ID of the known target, indicating that the target individual is a new target if the detected target individual is not successfully matched, and giving a new digital ID.
  5. 5. The machine vision-based intelligent inspection method of a farm according to claim 1, wherein obtaining abnormal body temperature of the target individual comprises: The thermal imaging image with the digital ID tag is input into a body temperature abnormality detection model to obtain the body temperature abnormality of the target individual, the body temperature abnormality detection model is obtained through training by using a training set, and the training set comprises an original thermal imaging image and a corresponding temperature tag.
  6. 6. The machine vision-based intelligent inspection method of a farm of claim 5, wherein training the body temperature anomaly detection model comprises: removing the image with serious noise pollution from the original thermal imaging image, and manually labeling the rest image, namely dividing targets in the rest image into abnormal body temperature and normal body temperature according to the display temperature of the thermal imaging image so as to form the training set; extracting the characteristics of the original thermal imaging images in the training set by adopting an improved full convolution neural network, and generating candidate object frames through a region recommendation network; Pooling and pixel alignment operations are carried out on the feature images processed by the region recommendation network, and the feature images processed by the pixel alignment operations are processed by using a convolution residual error module of the improved full convolution neural network, so that a region of interest is obtained; And carrying out fusion processing on the region of interest, then respectively outputting classification information, bounding box regression information and mask information according to three channels, and training a neural network by adopting a multi-task loss function.
  7. 7. The machine vision-based intelligent inspection method of a farm of claim 6, wherein improving the full convolutional neural network comprises: and introducing a full convolution neural network serving as a ResNet main network of the depth residual network into the full convolution neural network, namely fusing the high-level middle-low resolution characteristic map processed by the ResNet main network into the characteristic map of the full convolution neural network after passing through an initial convolution layer and a maximum pooling layer, namely the low-level middle-high resolution characteristic map, so that the characteristic map has stronger semantics.
  8. 8. The machine vision-based intelligent inspection method of a farm of claim 6, wherein generating candidate object frames through the area recommendation network comprises: Setting detection frames with different sizes and length-width ratios, namely anchor boxes, for each pixel of the feature map through the regional recommendation network; And allocating a binary label to each anchor box for distinguishing the foreground from the background, and carrying out regression operation on the predicted boundary box, so that the anchor boxes are mapped to obtain regression windows close to the real labeling boxes, and the candidate object frames are generated.

Description

Intelligent inspection method for farm based on machine vision Technical Field The invention relates to the technical field of intelligent inspection, in particular to an intelligent inspection method for a farm based on machine vision. Background With the increase of global population and the improvement of living standard, the demand for animal proteins is continuously increased, and the modern breeding industry is facing the opportunity of rapid development. The scale and the culture density of the farm are also improved obviously, which not only puts higher demands on the production efficiency of the cultured animals, but also puts serious challenges on the health management of the animals. Good animal health management is a key for guaranteeing the breeding benefit and animal welfare, and an effective inspection means is an important link for realizing health management. Limitations of the existing inspection mode: In the traditional cultivation mode, the inspection of the cultivation farm is mainly finished manually. Workers need to enter the breeding area regularly, and judge the health condition of the animals by observing the behaviors, the appearance and the environmental conditions of the animals by naked eyes. However, this manual inspection approach has a number of problems: The efficiency is low, a great deal of time and labor are consumed for manual inspection, and particularly in a large-scale farm, the timely and comprehensive inspection of all areas is difficult to realize. Workers need to shuttle back and forth between different cultivation areas, so that the efficiency is low, and important information is easy to miss due to fatigue. The subjectivity is strong, and the health condition of the animal is easily influenced by subjective factors, and the accuracy and consistency are lacking. The experience and judgment criteria may be different for different staff members, resulting in deviations in the assessment of animal health. The risk is high that animals may be disturbed by entering the farm, affecting their normal life, and even possibly inducing stress reactions of the animals, resulting in health problems. Furthermore, manual inspection may also increase the risk of disease transmission, especially when the animal is suffering from an infectious disease. The method is difficult to monitor in real time, the health data of animals in the farm cannot be obtained in real time, and abnormal conditions are difficult to discover in time and measures are taken. Once an animal has a health problem, it often takes a long time to find it, which may lead to a worsening condition or even spread to other animals. In recent years, with the rapid development of machine vision technology, the application of the machine vision technology in the inspection of farms is gradually paid attention to. The machine vision technology collects image and video data through equipment such as cameras, sensors and the like, analyzes and processes the data by utilizing a computer algorithm, and can realize automatic and intelligent inspection of a farm. However, existing machine vision inspection methods still have some limitations: The data association is insufficient, namely when a target individual is detected, the thermal imaging image cannot be effectively associated with the target in the inspection video, so that the information is fragmented, and the health condition of the target individual is difficult to comprehensively evaluate. For example, when detecting an animal abnormal body temperature, only the thermographic image is relied upon, and the behavior and appearance characteristics of the animal in the video data are ignored, which may lead to erroneous decisions. The target recognition accuracy is limited, namely, in a complex farm environment, the appearance characteristics of target individuals can be influenced by factors such as illumination, shielding, animal fur color similarity and the like, and the traditional target recognition method is difficult to accurately recognize and distinguish the target individuals. For example, when animals are shielded from each other or in low light conditions, the recognition accuracy may be greatly reduced. The abnormality detection capability is insufficient, namely, for the abnormal body temperature detection of a target individual, the accuracy and the reliability of the existing method are required to be improved, and the high requirements of a farm on animal health monitoring are difficult to meet. The existing body temperature detection model is sensitive to noise, and cannot effectively process complex backgrounds in images, so that the false alarm rate is high. The existing machine vision inspection system is lack of intelligent processing, and most of the existing machine vision inspection system only can provide simple image acquisition and preliminary analysis functions and lacks of deep mining and intelligent processing capacity for data. For exa