CN110866561A - Plastic bottle color sorting method based on image recognition - Google Patents

Abstract

The invention provides a plastic bottle color sorting method based on image recognition, which comprises the following steps: and S1, collecting the plastic bottle image, and dividing the image into a training image data set, a verification image data set and an image data set to be detected according to the proportion after carrying out denoising, averaging, defogging and cutting operations on the image. S2, establishing a deep learning network model for plastic bottle color sorting, inputting the training image data set into the network model for training, and determining the deep learning network model and model parameters. And S3, inputting the verification image data set into the determined deep learning network model, and checking the deep learning network model. And S4, inputting the image data set to be detected into the deep learning network model to obtain the plastic bottle color classification result in the image data set to be detected. The invention utilizes the deep learning image detection and classification to carry out real-time sorting on the colors of the plastic bottles, and can realize the plastic bottle color sorting result with high efficiency and strong accuracy.

Description

A Color Sorting Method of Plastic Bottles Based on Image Recognition

technical field

The invention relates to the technical field of image processing, in particular to a color sorting method for plastic bottles based on image recognition.

Background technique

With the development of the times and the improvement of people's living standards, the production and consumption of beverages continue to grow. For the disposal of plastic bottles, most of the current methods are burial and burning. Because the natural degradation time of plastic bottles is very long, and some harmful substances will be produced during the degradation process; the burning method will produce a large amount of harmful gases and cause serious environmental damage Pollution will also endanger people's health and cause waste of resources. Therefore, the country implements the recycling of plastic bottles. In order to ensure the quality of recycled materials, plastic bottles are classified by color. At present, the color classification of plastic bottles in my country mainly relies on manual sorting, and the efficiency is not high. It is a labor-intensive, time-consuming and cost-intensive work. The use of deep learning methods has greatly improved the work efficiency of color detection and classification of plastic bottles, and also saved a lot of labor and costs.

The current color recognition methods include: segmenting the image through edge detection and image binarization, and using the color difference relationship between the color in other areas after segmentation and the color in the reference area to determine the color classification; The RGB color image is converted into HSV color space, and the color is classified according to the H and V values. Existing methods are not suitable for experiments with large datasets, and the classification speed and accuracy are not high.

SUMMARY OF THE INVENTION

The invention provides a plastic bottle color sorting method with high accuracy, good rapidity and strong practicability, which is suitable for changing the color of plastic bottles in different environments.

A color sorting method for plastic bottles based on image recognition, comprising:

S1, collect an image of a plastic bottle, perform denoising, averaging, dehazing and cropping operations on the image, and divide the image into a training image data set, a verification image data set and an image data set to be tested in proportion;

S2, establish a deep learning network model for color sorting of plastic bottles, input the training image data set into the network model for training, and determine the deep learning network model and model parameters;

S3, input the verification image data set into the determined deep learning network model to test the deep learning network model;

S4, the image data set to be tested is input into the deep learning network model, and the color classification result of the plastic bottle in the image data set to be tested is obtained.

Further, the S1 includes the following steps:

S11, using an industrial camera to collect images of plastic bottles of different colors to form a first data set;

S12, performing Gaussian filtering and denoising on the images in the first data set;

S13, performing grayscale averaging and dehazing operations on the image processed in S12;

S14, the graphics processed in S13 are cropped according to the size of 224*224, and the cropped image is stored in the first data set, and the original image is replaced to form a second data set;

S15, classify the plastic bottles according to red, green, blue, purple, yellow, black, white and transparent colors for the images of the second data set in S14, label them, and perform operations such as rotating, mirroring, and increasing contrast in different directions. , and divide the second data set according to the ratio of 8:1:1 to obtain a preprocessed training image data set, a verification image data set and an image data set to be tested.

Further, the following steps are included in the S2:

S21, determine the convolution layer and the pooling layer of the input image;

S22, the feature extraction of the color of the plastic bottle is realized through the residual network;

S23, input the extracted feature map to the global average pooling layer, and then input the pooled features to the fully connected layer for plastic bottle color classification;

S24, set the parameters in the training algorithm, including the learning rate, the number of training times and the number of iterations;

S25, input the training image data set into the established network model, and determine the deep learning network model and model parameters.

Further, the convolution layer in S21 includes a first convolution layer and a second convolution layer, the size of the convolution kernel in the first convolution layer is 3*3, and the stride is 2; the second convolution layer The size of the convolution kernel in the convolutional layer is 3*3, and the stride is 1.

Further, the residual network described in S22 is provided with four residual blocks, and the residual blocks are provided with a third convolutional layer, a fourth convolutional layer and a fifth convolutional layer.

Further, the output calculation formula of the residual block is as follows:

Y=relu ₃ ([relu ₂ (relu ₁ (X,W ₁ ),W ₂ )]*W ₃ +X*Wt)

Among them, X is the input, Y is the output, W ₁ , W ₂ , and W ₃ are the weight matrices of each convolutional layer, relu ₁ is the excitation function output by the third-layer convolutional layer, and relu ₂ is the fourth-layer convolutional layer. The output excitation function, relu ₃ is the excitation function output by the residual block, and Wt is the linear transformation matrix.

Further, the convolution kernel of the third convolutional layer is 1*1, the stride is 2, the convolutional kernel of the fourth convolutional layer is 3*3, the stride is 1, and the volume of the fifth convolutional layer is The product kernel is 1*1, and the step size is 1.

Further, the S25 includes the following steps:

S251, determine whether the set number of training times is reached, and if so, obtain the result, otherwise repeat S251;

S252, determine whether the set number of iterations is reached, if so, the training ends, otherwise repeat S251;

S253, determine the deep learning network model and model parameters.

The beneficial effects of the present invention are:

1. Perform Gaussian filtering and denoising on the image, grayscale averaging, and dehazing, which overcomes the problem of interference to the image under different lighting conditions and improves the quality of the data set;

2. The image input uses two 3*3 convolution kernels (the first convolution kernel has a stride of 2, and the second convolution kernel has a stride of 1) instead of the 7*7 convolution kernel with a stride of 2, increasing the The number of layers of the network is increased, and the excitation function is added between the layers, thereby enhancing the nonlinear expression ability of the network and reducing the number of parameters;

3. For the residual block, the method of first reducing the network dimension and then restoring the network dimension also reduces the amount of computation.

Description of drawings

Fig. 1 is the overall flow chart of the present invention;

Fig. 2 is the concrete flow chart of the present invention S1;

Fig. 3 is the concrete flow chart of the present invention S2;

FIG. 4 is a structural diagram of a deep learning network model established by the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present invention.

1 to 4, a color sorting method for plastic bottles based on image recognition includes the following steps:

S1, collect an image of a plastic bottle, perform denoising, averaging, dehazing, and cropping operations on the image, and divide the image into a training image dataset, a validation image dataset, and a test image dataset in proportion. Specific steps are as follows:

S11, using an industrial camera to collect images of plastic bottles of different colors to form a first data set;

S12, performing Gaussian filtering and denoising on the images in the first data set;

S13 , performing a grayscale averaging operation on the image processed in S12. The specific method is as follows: dividing the color image into three independent single channels, then performing grayscale histogram equalization on each single channel, and then equalizing the equalized image The image is merged into one channel; then the image is dehazed to restore it to a color image;

S14, trim the graph processed in S13 according to the size of 224*224, store the trimmed image in the first data set, replace the original image, and form the second data set;

S15, classify the plastic bottles according to red, green, blue, purple, yellow, black, white and transparent colors for the images of the second data set in S14, label them, and perform operations such as rotating, mirroring, and increasing contrast in different directions. , and divide the second data set according to the ratio of 8:1:1 to obtain a preprocessed training image data set, a verification image data set and an image data set to be tested.

S2, establish a deep learning network model for color sorting of plastic bottles, input the training image data set into the network model for training, and determine the deep learning network model and model parameters. Specific steps are as follows:

S21, determine the convolution layer and the pooling layer of the input image. In order to enhance the nonlinear expression ability of the network, the convolution layer is set as the first convolution layer and the second convolution layer, the size of the convolution kernel in the first convolution layer is 3*3, and the stride is 2; The size of the convolution kernel in the second convolutional layer is 3*3, and the stride is 1.

The image processing in S21 is: the input image is 224*224*3, after the first convolution layer, the size of the convolution kernel is 3*3, the stride is 2, and 64 features are output, and the image size is 112*112; The second layer of convolutional layer, the convolution kernel size is 3*3, the stride is 1, the output is 64 features, and the image size is 112*112; The size is 56*56.

S22, the feature extraction of the color of the plastic bottle is realized through the residual network. In the residual network, in order to improve the accuracy and reduce the amount of calculation, there are four residual blocks, denoted as RB1, RB2, RB3, and RB4. There are three convolutional layers in each residual block, that is, the network dimension of the input layer is reduced first, then the features are learned, and finally the network dimension is restored. Among them, the three convolutional layers are the third convolutional layer, the fourth convolutional layer and the fifth convolutional layer respectively. The convolution kernel of the third convolutional layer is 1*1, the stride is 2, the convolutional kernel of the fourth convolutional layer is 3*3, the stride is 1, and the convolutional kernel of the fifth convolutional layer is 1* 1, the step size is 1.

The output of the residual block is calculated as follows:

Assuming that the input of the residual block is X, the output is Y, the weight matrices of each convolutional layer are W ₁ , W ₂ , W ₃ , the excitation function output by the third convolutional layer is relu ₁ , and the fourth convolutional layer The activation function output by the layer is relu ₂ , and the activation function output by the residual block is relu ₃ . In addition, in order to maintain the consistency of the input and output dimensions, assuming that the linear transformation matrix is Wt, the output:

Y=relu ₃ ([relu ₂ (relu ₁ (X,W ₁ ),W ₂ )]*W ₃ +X*Wt)

The image processing in S22 is: after the residual block RB1, output 256 features, the image size is 56*56; after the residual block RB2, output 512 features, the image size is 28*28; after the residual block RB3, output 1024 features, the image size is 14*14; after the residual block RB4, 2048 features are output, and the image size is 7*7.

S23, input the extracted feature map to the global average pooling layer, and then input the pooled features to the fully connected layer for plastic bottle color classification. The global average pooling operation is performed on the 2048 7*7 features output by the RB4 residual block, and 2048 1*1 pooling features are output, and then the pooled features are input to the fully connected layer, and finally the classification result is obtained.

S24, set the parameters in the training algorithm, including the learning rate, the number of training times, and the number of iterations.

S25, input the training image data set into the established network model, and determine the deep learning network model and model parameters. Specific steps are as follows:

S251, determine whether the set number of training times is reached, and if so, obtain the result, otherwise repeat S251;

S252, determine whether the set number of iterations is reached, if so, the training ends, otherwise repeat S251;

S253, determine the deep learning network model and model parameters.

S3, the verification image data set is input into the determined deep learning network model, and the deep learning network model is tested.

S4, the image data set to be tested is input into the deep learning network model, and the color classification result of the plastic bottle in the image data set to be tested is obtained.

The above description is only the preferred embodiment of the present invention, but the present invention should not be limited to the content disclosed in the embodiment and the accompanying drawings, so any equivalents or modifications accomplished without departing from the spirit disclosed in the present invention are all fall within the protection scope of the present invention.

Claims (8)

1. a plastic bottle color sorting method based on image recognition, is characterized in that, comprises: S1, collect an image of a plastic bottle, perform denoising, averaging, dehazing and cropping operations on the image, and divide the image into a training image data set, a verification image data set and an image data set to be tested in proportion; S2, establish a deep learning network model for color sorting of plastic bottles, input the training image data set into the network model for training, and determine the deep learning network model and model parameters; S3, input the verification image data set into the determined deep learning network model to test the deep learning network model; S4, the image data set to be tested is input into the deep learning network model, and the color classification result of the plastic bottle in the image data set to be tested is obtained. 2. a kind of plastic bottle color sorting method based on image recognition as claimed in claim 1, is characterized in that, comprises the following steps in described S1; S11, using an industrial camera to collect images of plastic bottles of different colors to form a first data set; S12, performing Gaussian filtering and denoising on the images in the first data set; S13, performing gray level averaging and dehazing operations on the image processed in S12; S14, the graphics processed in S13 are cropped according to the size of 224*224, and the cropped image is stored in the first data set, and the original image is replaced to form a second data set; S15, classify the plastic bottles according to red, green, blue, purple, yellow, black, white and transparent colors for the images of the second data set in S14, label them, and perform operations such as rotation in different directions, mirroring, increasing contrast, etc. , and divide the second data set according to the ratio of 8:1:1 to obtain a preprocessed training image data set, a verification image data set and an image data set to be tested. 3. a kind of plastic bottle color sorting method based on image recognition as claimed in claim 2, is characterized in that, comprises the following steps in described S2; S21, determine the convolution layer and the pooling layer of the input image; S22, the feature extraction of the color of the plastic bottle is realized through the residual network; S23, input the extracted feature map to the global average pooling layer, and then input the pooled features to the fully connected layer for plastic bottle color classification; S24, set the parameters in the training algorithm, including the learning rate, the number of training times and the number of iterations; S25, input the training image data set into the established network model, and determine the deep learning network model and model parameters. 4. a kind of plastic bottle color sorting method based on image recognition as claimed in claim 3 is characterized in that, the convolutional layer in described S21 comprises the first convolutional layer and the second convolutional layer, the said first convolutional layer. The size of the convolution kernel in the first convolutional layer is 3*3, and the stride is 2; the size of the convolution kernel in the second convolutional layer is 3*3, and the stride is 1. 5. The color sorting method for plastic bottles based on image recognition according to claim 3, wherein the residual network described in S22 is provided with four residual blocks, and the residual blocks are Set the third convolutional layer, the fourth convolutional layer and the fifth convolutional layer. 6. a kind of plastic bottle color sorting method based on image recognition as claimed in claim 5, is characterized in that, the output calculation formula of described residual block is as follows: Y=relu ₃ ([relu ₂ (relu ₁ (X,W ₁ ),W ₂ )]*W ₃ +X*Wt) Among them, X is the input, Y is the output, W ₁ , W ₂ , and W ₃ are the weight matrices of each convolutional layer, relu ₁ is the excitation function output by the third convolutional layer, and relu ₂ is the output of the fourth convolutional layer. The excitation function, relu ₃ is the excitation function output by the residual block, and Wt is the linear transformation matrix. 7. The color sorting method for plastic bottles based on image recognition according to claim 4, wherein the convolution kernel of the third convolution layer is 1*1, the step size is 2, and the fourth convolution layer The convolution kernel of the layer is 3*3 with a stride of 1, and the convolution kernel of the fifth convolutional layer is 1*1 with a stride of 1. 8. a kind of plastic bottle color sorting method based on image recognition as claimed in claim 3, is characterized in that, comprises the following steps in described S25: S251, determine whether the set number of training times is reached, and if so, obtain the result, otherwise repeat S251; S252, determine whether the set number of iterations is reached, if so, the training ends, otherwise repeat S251; S253, determine the deep learning network model and model parameters.

Images