CN-114821159-B - Identification method and system for classifying and recycling beverage bottles

CN114821159BCN 114821159 BCN114821159 BCN 114821159BCN-114821159-B

Abstract

The invention provides an identification method for classifying and recycling beverage bottles, which comprises the steps of collecting original images of the beverage bottles, sequentially carrying out pretreatment, labeling and data enhancement on the original images, dividing the treated original images into a training set and a testing set, carrying out fine adjustment stage parameter training on an instance segmentation model by using the training set and the testing set to obtain a trained instance segmentation model, predicting the original images by using the trained instance segmentation model to obtain the outline of each beverage bottle, carrying out matting treatment on the original images of each beverage bottle according to the outline of each beverage bottle to obtain characteristic images of each beverage bottle, extracting the average value of each characteristic on each characteristic image, and training a machine learning model by taking the average value as an input characteristic to obtain the trained machine learning model. Therefore, the beverage bottles can be accurately classified and recycled, and the problem that beverage bottles with similar colors are easy to confuse and identify in the recycling process is solved.

Inventors

JIANG FENGFENG
YANG JIANHONG
FANG HUAIYING
WANG ZHIFENG
YANG TIANCHENG
Ji Tianchen
XIE YIBIN
WANG ZHENG

Assignees

厦门陆海环保股份有限公司
漳州市陆海环保产业开发有限公司

Dates

Publication Date: 20260505
Application Date: 20220407

Claims (7)

1. The identification method for classifying and recycling the beverage bottles is characterized by comprising the following steps of: s100, collecting an original image of the beverage bottle; S200, preprocessing the acquired original image; s300, labeling the preprocessed original image; S400, carrying out data enhancement processing on the marked original image; S500, dividing the original image subjected to data enhancement processing into a training set and a testing set, loading an instance segmentation model based on a migration learning mode, performing fine tuning stage parameter training on the instance segmentation model by using the training set and the testing set, and evaluating the instance segmentation model by taking an average precision mean value as an index to obtain a trained instance segmentation model; S600, predicting the original image by using the trained example segmentation model to obtain the outline of each beverage bottle, then carrying out image matting processing on the original image of each beverage bottle according to the outline of each beverage bottle to obtain the characteristic image of each beverage bottle, and storing the characteristic image; S700, extracting an average value of each feature on each feature image, and training a machine learning model by taking the average value as an input feature; s800, evaluating the machine learning model to obtain a trained machine learning model; In the step S700, the average value of the three colors of red, green and blue on each feature image is extracted, then the format of the feature image is converted into HSV format, the average value of the three values of hue, saturation and brightness is calculated, and the average value of the six values of red, green, blue, hue, saturation and brightness is used as the input feature to train the machine learning model; In the step S500, the number ratio of the training set to the test set is 7:3, the loaded instance segmentation model adopts a Mask R-CNN architecture, then pre-training weights of an ImageNet data set are used as initial training parameters, fine tuning stage parameter training is performed, after training, an instance segmentation model with the highest average precision mean value is selected as a final used instance segmentation model, and the fine tuning stage parameter training refers to training mode of only training a full-connection layer by fixing weights of a feature extraction layer so as to obtain the instance segmentation model.
2. The method for sorting and recycling beverage bottles as set forth in claim 1, wherein after step S800 is completed, the method further comprises the steps of: s900, the trained instance segmentation model and the trained machine learning model are exported, and then the trained instance segmentation model and the trained machine learning model are deployed.
3. The method for sorting and recycling beverage bottles as set forth in claim 1, wherein in step S100, an original image of the de-labeled beverage bottles is acquired by an industrial camera.
4. The method for sorting and recycling beverage bottles as set forth in claim 1, wherein in said step S200, said preprocessing includes an image cropping process and an image scaling process.
5. The method according to claim 1, wherein in the step S300, the labeling process is a semantic labeling process, and the labeling format is a COCO format.
6. The method for sorting recycling of beverage bottles as set forth in claim 1, wherein in said step S400, said data enhancement process includes image inversion and image duplication.
7. The method according to claim 1, wherein in step S800, the machine learning model is evaluated with respect to recovery rate and purity.

Description

Identification method and system for classifying and recycling beverage bottles Technical Field The invention relates to the technical field of beverage bottle classification recovery, in particular to an identification method and an identification system for beverage bottle classification recovery. Background The beverage bottle is one of the most widely used plastic products in life, most of the beverage bottle is made of PET, and the waste beverage bottle is recycled, so that the pollution of the beverage bottle to the environment and the resource waste can be effectively reduced, and the recycling of resources is promoted. At present, beverage bottles are required to be classified according to colors in the recycling process of the beverage bottles, so that plastics with different colors can be conveniently regenerated, and recycling benefits are maximized. In the traditional beverage bottle recycling process, the beverage bottles are needed to be manually selected from the mixed household garbage, the workload of workers in the sorting process is large, the working environment is bad, and the large-scale recycling of the waste beverage bottles is not facilitated. At present, an automatic sorting device based on beverage bottle color images is available, and the beverage bottles are identified by a deep learning method, so that a large amount of manually marked training sample data is required, the problem of low detection precision exists, and particularly, the sorting effect on objects (such as transparent objects and light blue objects) with similar beverage bottle colors is poor. The method for identifying colors is used for the beverage bottles after label removal, and the outlines of the beverage bottles are required to be extracted by threshold segmentation to position the beverage bottles. The actual sorting working condition is influenced by factors such as complex background and light change, so that the outline of the beverage bottle cannot be extracted well due to threshold segmentation, and the accuracy of color identification is further influenced. Therefore, how to accurately classify and recycle beverage bottles has become a technical problem to be solved urgently by those skilled in the art. Disclosure of Invention The invention provides an identification method for classifying and recycling beverage bottles, which comprises the following steps of S100, collecting original images of the beverage bottles; the method comprises the steps of S200, preprocessing an acquired original image, S300, marking the preprocessed original image, S400, data enhancement processing of the marked original image, S500, dividing the original image subjected to the data enhancement processing into a training set and a testing set, loading an instance segmentation model based on a migration learning mode, performing fine tuning stage parameter training on the instance segmentation model by using the training set and the testing set, evaluating the instance segmentation model by using an average precision mean value as an index to obtain a trained instance segmentation model, S600, predicting the original image by using the trained instance segmentation model to obtain the outline of each beverage bottle, performing matting processing of the original image of each beverage bottle according to the outline of each beverage bottle to obtain the characteristic image of each beverage bottle, and storing the characteristic image, S700, extracting the average value of each characteristic on each characteristic image, and training a machine learning model by using the average value as an input characteristic, and S800, evaluating the machine learning model to obtain the trained machine learning model. In one embodiment, after step S800 is completed, the method further comprises the step S900 of deriving the trained instance segmentation model and the trained machine learning model, and deploying the trained instance segmentation model and the trained machine learning model In one embodiment, in the step S100, an original image of the off-label beverage bottle is acquired by an industrial camera. In an embodiment, in the step S200, the preprocessing includes an image cropping process and an image scaling process. In an embodiment, in the step S300, the labeling process is a semantic labeling process, and the labeling format is a COCO format. In an embodiment, in the step S400, the data enhancement process includes image flipping and image copying. In an embodiment, in the step S500, the number ratio of the training set to the test set is 7:3, and the loaded instance segmentation model adopts a Mask R-CNN architecture. In one embodiment, in the step S700, the average value of the three colors of red, green and blue on each of the feature images is extracted, then the format of the feature images is converted into HSV format, the average value of the three values of hue, saturation and brightness is calculated, and the a