Construction method of dairy cow nipple detection convolutional neural network model

Abstract

The invention discloses a construction method of a dairy cow nipple detection convolutional neural network model, the construction method comprises the following steps: acquiring at least one dairy cow image of at least one dairy cow, wherein the dairy cow image comprises a dairy cow nipple; obtaining an annotation image corresponding to each dairy cow image, with the annotation image having at least one target frame, and the target frame being used for annotating the position of the dairy cow nipple in the dairy cow image; dividing the dairy cow image and the corresponding annotation image into a training set, a verification set and a test set according to a preset proportion; constructing a dairy cow nipple detection convolutional neural network model based on the dairy cow images in thetraining set and the corresponding annotation images; and optimizing and testing the dairy cow nipple detection convolutional neural network model parameters based on the dairy cow images in the verification set and the test set and the corresponding annotation images.

Description

Technical field [0001] The present invention relates to the computer field, and further relates to a milk head detection convolutional neural network model and its construction method and its construction system. Background technique [0002] With the development of science and technology, the development of automation in agriculture and animal husbandry has increasingly become a current demand. The development of automation can reduce a lot of manpower and time costs, while facilitating human management, such as automatic spraying of pesticides on farms, automatic feeding devices for pastures, etc. Wait. Among the many animal husbandry production labor, milking cows has always been a job with high labor intensity and higher requirements for operators. At present, some researches have designed automatic milking devices, but these devices still require Manual operation, complicated operation, and low work efficiency. [0003] It should be understood by those skilled in the art tha...

AI Technical Summary

Problems solved by technology

The selection of regions usually uses an exhaustive strategy, adopts a sliding window, and sets different sizes to traverse the image with different aspect ratios. The time complexity is high. Secondly, the feature extraction used is artificially designed. Due to the variety of shapes Sex, illumination change diversity, background diversity, etc. make the robustness poor

[0004] With the development of convolutional neural network architecture, object detection and recognition based on deep learning has become the mainstream method. One type contains object boundary area suggestions, such as R-CNN, Fast R-CNN, Faster R-CNN and other models. The class model uses the enumeration method to pre-assume the target candidate area, and gradually fine-tunes and optimizes the target position, and finally realizes its classification and recognition, but the detection speed of this type of method is slow

The other type is to directly generate boundary areas, such as one-stage detection models such as YOLO and SSD. This type of model can simultaneously predict boundaries and classify and identify targets when predicting targets. The detection speed is significantly improved, but there are often The problem of low detection accuracy such as missed detection

On the other hand, the deep learning framework needs to learn too many parameters, the calculation speed is relatively slow, and overfitting may occur, so there are still many areas for improvement in target detection.

Images