Method for identifying degradability of solid waste on-site compost

Abstract

The invention discloses a method for identifying degradability of solid waste field compost, which comprises the following steps of: 1, capturing an original solid waste image through a camera, cutting an ROI (Region of Interest) of a to-be-processed target from the original image, and cutting the ROI into a plurality of small block images to generate a test set; 2, predicting the degradability of a test set by using a trained lightweight convolutional neural network (Lite-CNN) model; 3, according to the solid waste degradability result predicted in the step 2, separation of non-degradable solid waste is guided, and when the degradability of the solid waste at an inlet reaches a required threshold value, the solid waste enters a composting reactor. According to the method, the degradability of the community household garbage can be effectively identified, and the accuracy is high.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method for identifying the biodegradability of solid waste in on-site composting. Background Technology

[0002] Solid waste refers to all solid or semi-solid materials generated by human activities that are no longer of use to their owners and are therefore discarded. Solid waste can be divided into two categories: biodegradable and non-biodegradable. Biodegradable waste refers to pollutants that can be degraded into harmless substances in the natural environment; these are materials suitable for composting, such as vegetable leaves, fruits, weeds, and food scraps. Non-biodegradable waste, on the other hand, refers to garbage that cannot be degraded into harmless waste, such as plastic film, printed packaging bags, and beverage bottles. Automatic identification of the biodegradability of solid waste in communities is one of the key prerequisites for preventing pollution from non-compostable waste.

[0003] Convolutional Neural Networks (CNNs) are a class of feedforward neural networks that incorporate convolutional computations and have deep structures. They are one of the representative algorithms in deep learning. The AlexNet neural network structure is a type of CNN. AlexNet deepens the network structure based on LeNet, learning richer and higher-dimensional image features. AlexNet's characteristics include: a deeper network structure; the use of stacked convolutional layers (convolutional layer + convolutional layer + pooling layer) to extract image features; Dropout to suppress overfitting; and ReLU replacing the previous sigmoid as the activation function. How to utilize CNNs to improve the accuracy of identifying the biodegradability of community household waste is a pressing research question. Summary of the Invention

[0004] In view of this, the present invention proposes a method for identifying the biodegradability of solid waste composting on-site. Based on the AlexNet neural network structure, a lightweight convolutional neural network (Lite-CNN) model that is easy to train and implement is established. Through experiments, the optimal combination of image cropping sizes of the training set and test set after image enhancement processing is found to improve the accuracy of identifying the biodegradability of community domestic waste.

[0005] On the one hand, this invention proposes a method for identifying the biodegradability of solid waste in on-site composting, comprising the following steps: Step 1: Capture the original image of solid waste using a camera, crop out the Region of Interest (ROI) of the target to be processed from the original image, and cut the ROI into multiple small square images to generate a test set; Step 2: Predict the degradability of the test set using the trained lightweight convolutional neural network (Lite-CNN) model; Step 3: Based on the predicted solid waste degradability results in Step 2, guide the separation of non-degradable solid waste. When the degradability of the solid waste at the inlet reaches the required threshold, allow the solid waste to enter the composting reactor.

[0006] In a further technical solution, step one specifically includes: cutting the region of interest (ROI) into small square images with a side length of 1 / 14 of the original side length, resulting in a number of small square images that are 196 times the number of the original image.

[0007] In a further technical solution, in step two, a lightweight convolutional neural network (Lite-CNN) model, easy to train and implement, is established based on the AlexNet convolutional neural network structure. The input image is adjusted to a color image with three RGB channels and a width × height pixel size of 48 × 48. This image with an input feature matrix of [48 × 48 × 3] is used as the input layer. After the input layer, one convolutional layer is deployed, followed by four convolutional layers and four max-pooling layers at intervals. Then, three fully connected layers and two dropout layers are connected, with the two dropout layers deployed between two adjacent fully connected layers. Finally, the softmax activation function is applied as the output layer. The output of this lightweight convolutional neural network (Lite-CNN) model consists of two values ​​from 0 to 1, predicting the probability of solid waste in the image being degradable or non-degradable. The layers after the input layer of the lightweight convolutional neural network (Lite-CNN) model are arranged as follows: The first convolutional layer has a kernel size of 1×1, a stride of 1, an input feature matrix of [48×48×32], an output feature matrix of [48×48×32], and 128 training parameters; A second convolutional layer is deployed after the first convolutional layer. The kernel size of the second convolutional layer is 5×5, the stride is 1, the input feature matrix is ​​[48×48×32], the output feature matrix is ​​[48×48×32], and the training parameters are 25632. After the second convolutional layer, the first max pooling layer is deployed. The kernel size of the first max pooling layer is 3×3, the stride is 2, the input feature matrix is ​​[48×48×32], and the output feature matrix is ​​[24×24×32]. After the first max pooling layer, a third convolutional layer is deployed. The kernel size of the third convolutional layer is 3×3, the stride is 1, the input feature matrix is ​​[24×24×32], the output feature matrix is ​​[24×24×32], and the training parameters are 9248. After the third convolutional layer, a second max pooling layer is deployed. The kernel size of the second max pooling layer is 3×3, the stride is 2, the input feature matrix is ​​[24×24×32], and the output feature matrix is ​​[12×12×32]. After the second max pooling layer, a fourth convolutional layer is deployed. The kernel size of the fourth convolutional layer is 3×3, the stride is 1, the input feature matrix is ​​[12×12×32], the output feature matrix is ​​[12×12×32], and the training parameters are 9248. After the fourth convolutional layer, a third max pooling layer is deployed. The kernel size of the third max pooling layer is 3×3, the stride is 2, the input feature matrix is ​​[12×12×32], and the output feature matrix is ​​[6×6×32]. After the third max pooling layer, a fifth convolutional layer is deployed. The kernel size of the fifth convolutional layer is 5×5, the stride is 1, the input feature matrix is ​​[6×6×32], the output feature matrix is ​​[6×6×64], and the training parameters are 51264. After the fifth convolutional layer, a fourth max pooling layer is deployed. The kernel size of the fourth max pooling layer is 3×3, the stride is 2, the input feature matrix is ​​[6×6×64], and the output feature matrix is ​​[3×3×64]. After the fourth max-pooling layer, a first fully connected layer with 2048 neurons and 1,181,696 training parameters is deployed. Following this, a first dropout layer with 2048 neurons is deployed. Next, a second fully connected layer with 1024 neurons and 2,098,176 training parameters is deployed. Then, a second dropout layer with 1024 neurons is deployed. Finally, a third fully connected layer with 2 neurons and 2050 training parameters is deployed. Finally, a softmax activation function is applied as the output layer, transforming the output of the neurons in the third fully connected layer into a probability output indicating whether solid waste is biodegradable or non-biodegradable.

[0008] In a further technical solution, in step two, the lightweight convolutional neural network (Lite-CNN) model is trained using TensorFlow software. TensorFlow is an open-source framework for machine learning and deep learning. The process of generating training and validation sets is as follows: images of non-degradable waste samples and images of degradable waste samples are collected. The collected sample images are then cut, expanded, and enhanced to generate training and validation set data. First, the collected sample images are adjusted to be images of the central area covered by solid waste. Then, the central area images are cut into small square sample images with a side length of 1 / 8 of the original side length. The squares are rotated at a random angle of ±45°, and translated and randomly flipped horizontally along the width and height directions at a random distance of 0.2 times the side length to generate training and validation set data.

[0009] The advantages of the solid waste on-site composting biodegradability identification method of the present invention compared with the prior art are as follows: 1. Based on the AlexNet neural network structure, the central region image is cut into small square sample images with a side length of 1 / 8 of the original side length. The sample images are rotated at a random angle of ±45°, translated at a random distance of 0.2 times the side length in the width and height directions, and randomly flipped horizontally to generate training and validation set data. A lightweight convolutional neural network (Lite-CNN) model that is easy to train and implement is established. The validation accuracy of this model reaches 98.48%.

[0010] 2. The Region of Interest (ROI) is divided into small square images with sides equal to 1 / 14 of the original side length to generate test set data. This test set, together with the training and validation sets in the Lightweight Convolutional Neural Network (Lite-CNN) model, forms the optimal combination, achieving an accuracy of 91.58% in identifying the biodegradability of community household waste. Attached Figure Description

[0011] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the lightweight convolutional neural network (Lite-CNN) model structure of the present invention; Figure 2 This is a schematic diagram illustrating the accuracy comparison of the lightweight convolutional neural network (Lite-CNN) model of the present invention. Figure 3 This is a schematic diagram comparing the test accuracy results of the lightweight convolutional neural network (Lite-CNN) model of the present invention; Figure 4 This is a schematic diagram comparing the test accuracy corresponding to the test image cutting size for the lightweight convolutional neural network (Lite-CNN) model of the present invention. Detailed Implementation

[0012] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0013] This invention preferably provides a method for identifying the biodegradability of solid waste in on-site composting, comprising the following steps: Step 1: Capture the original image of solid waste using a camera, crop out the Region of Interest (ROI) of the target to be processed from the original image, and cut the ROI into small square images with a side length of 1 / 14 of the original side length to generate a test set. The number of small square images generated is 196 times that of the original image.

[0014] The camera has an image resolution of 1280×720, a minimum viewing angle of 50°, and is illuminated by a 300 lx LED light source at a temperature of 4000K. It is mounted at the same height and angle as the door. The camera is fixed to the inside of the door, and its imaging area completely covers the entrance to the composting unit. The region of interest (ROI) is approximately 32 cm × 32 cm.

[0015] Step 2: Use the trained Lightweight Convolutional Neural Network (Lite-CNN) model to predict the degradability of the test set.

[0016] Step 3: Based on the predicted solid waste degradability results from Step 2, guide the separation of non-degradable solid waste. When the degradability of the solid waste at the inlet reaches the required threshold, allow the solid waste to enter the composting reactor.

[0017] like Figure 1 As shown, a lightweight convolutional neural network (Lite-CNN) model that is easy to train and implement is established based on the convolutional neural network structure of AlexNet. The input image is adjusted to a color image with three RGB channels and a width × height pixel size of 48×48. The image of the input feature matrix [48×48×3] is used as the input of the input layer. After the input layer (input data), one convolutional layer (Conv1) is deployed, followed by four convolutional layers (Conv2, Conv3, Conv4, Conv5) and four max pooling layers (not shown). Then, three fully connected layers (FC6, FC7, FC8) and two dropout layers (not shown) are connected. The two dropout layers are deployed between two adjacent fully connected layers (FC6, FC7, FC8). The dropout layers reduce the complexity of the model by randomly stopping a batch of neurons each time, thus making the model less prone to overfitting. Finally, the softmax activation function is applied as the output layer; the output of this lightweight convolutional neural network (Lite-CNN) model consists of two values ​​from 0 to 1, predicting the probability of solid waste in the image being biodegradable or non-biodegradable. The layers following the input layer of the lightweight convolutional neural network (Lite-CNN) model are arranged as follows: like Figure 1 and Figure 2As shown, the kernel size of the first convolutional layer is 1×1, the stride is 1, the input feature matrix is ​​[48×48×32], the output feature matrix is ​​[48×48×32], and the training parameters are 128. A second convolutional layer is deployed after the first convolutional layer. The kernel size of the second convolutional layer is 5×5, the stride is 1, the input feature matrix is ​​[48×48×32], the output feature matrix is ​​[48×48×32], and the training parameters are 25632. After the second convolutional layer, the first max pooling layer is deployed. The kernel size of the first max pooling layer is 3×3, the stride is 2, the input feature matrix is ​​[48×48×32], and the output feature matrix is ​​[24×24×32]. After the first max pooling layer, a third convolutional layer is deployed. The kernel size of the third convolutional layer is 3×3, the stride is 1, the input feature matrix is ​​[24×24×32], the output feature matrix is ​​[24×24×32], and the training parameters are 9248. After the third convolutional layer, a second max pooling layer is deployed. The kernel size of the second max pooling layer is 3×3, the stride is 2, the input feature matrix is ​​[24×24×32], and the output feature matrix is ​​[12×12×32]. After the second max pooling layer, a fourth convolutional layer is deployed. The kernel size of the fourth convolutional layer is 3×3, the stride is 1, the input feature matrix is ​​[12×12×32], the output feature matrix is ​​[12×12×32], and the training parameters are 9248. After the fourth convolutional layer, a third max pooling layer is deployed. The kernel size of the third max pooling layer is 3×3, the stride is 2, the input feature matrix is ​​[12×12×32], and the output feature matrix is ​​[6×6×32]. After the third max pooling layer, a fifth convolutional layer is deployed. The kernel size of the fifth convolutional layer is 5×5, the stride is 1, the input feature matrix is ​​[6×6×32], the output feature matrix is ​​[6×6×64], and the training parameters are 51264. After the fifth convolutional layer, a fourth max pooling layer is deployed. The kernel size of the fourth max pooling layer is 3×3, the stride is 2, the input feature matrix is ​​[6×6×64], and the output feature matrix is ​​[3×3×64]. After the fourth max-pooling layer, a first fully connected layer with 2048 neurons and 1,181,696 training parameters is deployed. Following this, a first dropout layer with 2048 neurons is deployed. Next, a second fully connected layer with 1024 neurons and 2,098,176 training parameters is deployed. Then, a second dropout layer with 1024 neurons is deployed. Finally, a third fully connected layer with 2 neurons and 2050 training parameters is deployed. Finally, a softmax activation function is applied as the output layer, transforming the output of the neurons in the third fully connected layer into a probability output indicating whether solid waste is biodegradable or non-biodegradable.

[0018] A lightweight convolutional neural network (Lite-CNN) model was trained using TensorFlow, an open-source framework for machine learning and deep learning. The computer hardware used for training the Lite-CNN model included an Intel i9-10920 CPU and an RTX TITAN graphics card. The hyperparameters were set as follows: the optimizer was stochastic gradient descent (SGD), the loss function was categorical cross-entropy, the initial learning rate was 0.001, the learning rate decay was 0.00001, the momentum was 0.9, the batch size was 32, the dropout coefficient was 0.5, and the number of steps per epoch was 500. SGD minimizes the objective function... The algorithm,

[0019] Among them, based on each training sample and tags In have Calculate the loss function in mini-batches of samples. Then update the parameters using the following equation. ,

[0020] in, It is the gradient operator, learning rate This determines the step size used in the process of reaching the minimum value.

[0021] In addition, to achieve better training performance, the input data is centrally normalized using the following formula:

[0022] in, and These are the mean and standard deviation of all pixels in a dataset. The dataset is randomly divided into a training set and a validation set in an 8:2 ratio. The generation process of the training set and validation set is as follows: Images of non-degradable and degradable waste samples were collected. The collected sample images were then cut, expanded, and enhanced to generate training and validation sets.

[0023] Under good sunlight conditions, images were acquired twice a day, in the morning and afternoon, avoiding direct sunlight to obtain stable light intensity. The acquired images have a resolution of 4096×3072 pixels and are approximately 40cm×30cm in size. A total of 144 original images were acquired for both non-biodegradable and biodegradable waste samples, with 72 images for each category. First, the collected sample images were resized to show the central area covered by solid waste, approximately 32cm×32cm. Then, the central area image was cropped into smaller square images with sides of 1 / 2, 1 / 4, 1 / 6, and 1 / 8 of the original side length. Therefore, the sizes of these smaller square images are approximately 16cm×16cm, 8cm×8cm, 5.3cm×5.3cm, and 4cm×4cm, respectively. After removing images with significant blurriness, four original training sets were established, named W2, W4, W6, and W8 according to their cropping size, with 576, 2296, 5156, and 9088 training images, respectively. Training convolutional neural networks requires a large amount of labeled data to avoid overfitting. Therefore, this invention employs image augmentation techniques to expand the dataset. Image augmentation operations include rotation at random angles of ±45°, translation along the width and height directions at random distances of 0.2 times the side length, and horizontal random flipping. These operations fill blank areas in the image with a mirrored pattern. Through these operations, the number of trainable images increases tenfold, establishing four augmented training sets, named W2A, W4A, W6A, and W8A, respectively. The number of images in these four augmented training sets is ten times that of the original training set.

[0024] Eight training models were generated by training a lightweight convolutional neural network (Lite-CNN) model using four augmented training sets and four original training sets, respectively. The training models from the first five training epochs were then evaluated on a validation set, and the average validation accuracy was recorded, along with the standard deviation of these results.

[0025] like Figure 3As shown, the model trained on the original W8 training set with the smallest cropped size achieved a highest validation accuracy of 97.57%, which is 0.85 percentage points higher than the result obtained by training the model on the original W2 training set. The results indicate that the model trained on the four augmented training sets significantly outperforms the model trained on the four original training sets, with accuracy improvements of 2.07, 1.08, 1.61, and 0.91 percentage points, respectively. The model trained using the W8A augmented training set achieved a highest validation accuracy of 98.48%.

[0026] like Figure 4 As shown, to evaluate the performance of the biodegradability identification system and find the optimal cut size prediction, experiments were conducted where the input ROI image was cut into small square images with side lengths of 1 / 2, 1 / 4, 1 / 6, 1 / 8, 1 / 10, 1 / 12, 1 / 14, and 1 / 16 of the original side length. These small square images were then predicted using four models trained on four augmented training sets: W2A, W4A, W6A, and W8A. Eight samples were collected for evaluation, including biodegradable waste (vegetable scraps and melon rinds) and non-biodegradable waste (plastic film, plastic bottles, cardboard, and waste paper). Each evaluation was conducted on a mixed sample, which consisted of at least two types of waste and included both biodegradable and non-biodegradable waste. The evaluation result was the average of three repeated trials for each sample. Figure 4 As can be seen, except for models trained using the W2A and W4A augmented training sets, which achieve an accuracy exceeding 80%, the accuracy is relatively low when predicting images with large cut sizes, such as 62.50%-78.10% when the cut size is 1 / 2 or 1 / 4 of the side length. Furthermore, it was found that the smaller the cut size, the higher the accuracy. The model trained using the W8A augmented training set achieves the highest accuracy of 91.58% when the cut size is 1 / 14. The model trained using the W2A augmented training set achieves an accuracy of 90.95% at the same cut size. The accuracy of the remaining models is all below 90%. These results indicate that in degradability recognition systems, good performance can be achieved when the training augmented training set and the cut size of the test images are appropriately selected.

[0027] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying the biodegradability of solid waste in on-site composting, characterized in that, Includes the following steps: Step 1: Capture the original image of solid waste using a camera, crop out the Region of Interest (ROI) of the target to be processed from the original image, and cut the ROI into multiple small square images to generate a test set; Step 2: Predict the degradability of the test set using the trained lightweight convolutional neural network (Lite-CNN) model; Step 3: Based on the predicted solid waste degradability results from Step 2, guide the separation of non-degradable solid waste. When the degradability of the solid waste at the inlet reaches the required threshold, allow the solid waste to enter the composting reactor.

2. The method for identifying the biodegradability of solid waste in on-site composting according to claim 1, characterized in that, Step one specifically includes: cutting the region of interest (ROI) into small square images with a side length of 1 / 14 of the original side length, resulting in a number of small square images that are 196 times the number of the original image.

3. The method for identifying the biodegradability of solid waste in on-site composting according to claim 1, characterized in that, In step two, a lightweight convolutional neural network (Lite-CNN) model, easy to train and implement, is established based on the AlexNet convolutional neural network architecture. The input image is adjusted to a color image with three RGB channels and a width × height pixel size of 48 × 48. This image with an input feature matrix of [48 × 48 × 3] is used as the input layer. After the input layer, one convolutional layer is deployed, followed by four convolutional layers and four max pooling layers at intervals. Then, three fully connected layers and two dropout layers are connected, with the two dropout layers deployed between two adjacent fully connected layers. Finally, a softmax activation function output layer is deployed. The output of this lightweight convolutional neural network (Lite-CNN) model consists of two values ​​from 0 to 1, which predict the probability of solid waste in the image being biodegradable or non-biodegradable. The layers after the input layer of the lightweight convolutional neural network (Lite-CNN) model are arranged as follows: The first convolutional layer has a kernel size of 1×1, a stride of 1, an input feature matrix of [48×48×32], an output feature matrix of [48×48×32], and 128 training parameters; A second convolutional layer is deployed after the first convolutional layer. The kernel size of the second convolutional layer is 5×5, the stride is 1, the input feature matrix is ​​[48×48×32], the output feature matrix is ​​[48×48×32], and the training parameters are 25632. After the second convolutional layer, the first max pooling layer is deployed. The kernel size of the first max pooling layer is 3×3, the stride is 2, the input feature matrix is ​​[48×48×32], and the output feature matrix is ​​[24×24×32]. After the first max pooling layer, a third convolutional layer is deployed. The kernel size of the third convolutional layer is 3×3, the stride is 1, the input feature matrix is ​​[24×24×32], the output feature matrix is ​​[24×24×32], and the training parameters are 9248. After the third convolutional layer, a second max pooling layer is deployed. The kernel size of the second max pooling layer is 3×3, the stride is 2, the input feature matrix is ​​[24×24×32], and the output feature matrix is ​​[12×12×32]. After the second max pooling layer, a fourth convolutional layer is deployed. The kernel size of the fourth convolutional layer is 3×3, the stride is 1, the input feature matrix is ​​[12×12×32], the output feature matrix is ​​[12×12×32], and the training parameters are 9248. After the fourth convolutional layer, a third max pooling layer is deployed. The kernel size of the third max pooling layer is 3×3, the stride is 2, the input feature matrix is ​​[12×12×32], and the output feature matrix is ​​[6×6×32]. After the third max pooling layer, a fifth convolutional layer is deployed. The kernel size of the fifth convolutional layer is 5×5, the stride is 1, the input feature matrix is ​​[6×6×32], the output feature matrix is ​​[6×6×64], and the training parameters are 51264. After the fifth convolutional layer, a fourth max pooling layer is deployed. The kernel size of the fourth max pooling layer is 3×3, the stride is 2, the input feature matrix is ​​[6×6×64], and the output feature matrix is ​​[3×3×64]. After the fourth max-pooling layer, a first fully connected layer with 2048 neurons and 1,181,696 training parameters is deployed. Following this, a first dropout layer with 2048 neurons is deployed. Next, a second fully connected layer with 1024 neurons and 2,098,176 training parameters is deployed. Then, a second dropout layer with 1024 neurons is deployed. Finally, a third fully connected layer with 2 neurons and 2050 training parameters is deployed. Finally, a softmax activation function is applied as the output layer, transforming the output of the neurons in the third fully connected layer into a probability output indicating whether solid waste is biodegradable or non-biodegradable.

4. The method for identifying the biodegradability of solid waste in on-site composting according to claim 3, characterized in that, In step two, the lightweight convolutional neural network (Lite-CNN) model is trained using TensorFlow software. TensorFlow is an open-source framework for machine learning and deep learning. The process of generating the training and validation sets is as follows: images of non-degradable waste samples and images of degradable waste samples are collected. The collected sample images are then cut, expanded, and enhanced to generate training and validation set data. First, the collected sample images are adjusted to be images of the central area covered by solid waste. Then, the central area images are cut into small square sample images with a side length of 1 / 8 of the original side length. The squares are rotated at random angles of ±45°, and translated and randomly flipped horizontally at random distances of 0.2 times the side length in the width and height directions to generate training and validation set data.

Images