Powder pneumatic conveying image processing method, device and storage medium

Abstract

The application relates to a powder pneumatic conveying image processing method and device and a storage medium, which are applied to the field of image processing and include the following steps: acquiring an image of a material accumulation profile of a feeding port and corresponding feeding text information, taking the image of the material accumulation profile of the feeding port as an original profile image; adopting an adaptive contrast enhancement fusion color correction algorithm to pretreat the original profile image and generating a target profile image; adding a preset number of Gaussian noises to the target profile image to generate a noise profile image; inputting the noise profile image and the corresponding feeding text information into a trained profile enhancement neural network model and outputting a clear profile image after Bayesian denoising. The application has the technical effect that a profile clear material accumulation image is obtained, and a dynamic change process of the material accumulation profile of the feeding port can be seen.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image data processing, in particular to a powder pneumatic conveying image processing method, device and storage medium. BACKGROUND

[0002] In the pharmaceutical, food, chemical, new energy and other industries, ton bag, small bag of powder or granular material in the feeding process, usually artificial feeding method, or using mechanical arm or other conveyors to feed the feeding port. Artificial feeding method is generally low efficiency, mechanical arm or other conveyors to feed the feeding port can solve the problem of low efficiency of artificial feeding, but the mechanical arm feeding may have the problem of difficult to find the blockage in time, which may further affect the efficiency and cause material waste. There are many conventional parameters for monitoring the working state of the feeding port, including feeding port level meter, flow, pipe internal gas pressure value, etc. Video image is usually used to monitor the working state of the feeding port; however, the video image captured by the camera will have powder dust, resulting in blurred and unclear video image captured by the camera.

[0003] At present, there are some image data processing methods for processing the image of the feeding port to obtain clear video image; however, after processing by the current image processing method, it is still difficult to clearly see the material accumulation profile of the feeding port through the image, so as to cause difficulty in judging whether the discharge port has serious accumulation trend based on the material accumulation profile of the feeding port, thereby causing blockage and affecting the feeding efficiency. SUMMARY

[0004] In order to help solve the problem of difficulty in clearly seeing the material accumulation profile of the feeding port through the image, causing difficulty in judging whether the discharge port has serious accumulation trend based on the material accumulation profile of the feeding port, causing blockage and affecting the feeding efficiency, the present application provides a powder pneumatic conveying image processing method, device and storage medium.

[0005] In the first aspect, the present application provides a powder pneumatic conveying image processing method, which adopts the following technical scheme: the method comprises:

[0006] Obtaining the image of the material accumulation profile of the feeding port and the corresponding feeding text information, and setting the image of the material accumulation profile of the feeding port as the original profile image;

[0007] Using an adaptive contrast enhancement fusion color correction algorithm to preprocess the original profile image, and generating a target profile image;

[0008] Adding a preset number of Gaussian noise to the target profile image to generate a noise profile image;

[0009] The noise profile image and the corresponding feeding text information are input into the trained profile enhancement neural network model, and a clear profile image is output after Bayesian denoising.

[0010] In a specific implementation, the original profile image is a color image, the original profile image is in RGB form, the original profile image is preprocessed by using the adaptive contrast enhancement fusion color correction algorithm, and a target profile image is generated, which includes:

[0011] The original profile image in RGB form is converted into an original profile image in HSI form, which includes an H channel, an S channel and an I channel, and the conversion includes:

[0012]

[0013] H = 360°-H (if B > G)

[0014]

[0015] Wherein, R, G, B represent the data of R channel, G channel and B channel of original profile image in RGB form respectively, H, S, I represent the data of H channel, S channel and I channel of original profile image in HSI form respectively after conversion;

[0016] The data of H channel, S channel and I channel are normalized respectively to generate normalized H array, normalized S array and normalized I array, and the calculation of data normalization of H channel, S channel and I channel includes:

[0017] S norm = S,

[0018] Wherein, H norm represents the normalized H array, S norm represents the normalized S array, I norm represents the normalized I array, I max represents the maximum value in I channel data;

[0019] The normalized I array, the normalized H array and the normalized S array are processed by using adaptive contrast enhancement algorithm, hue correction algorithm and saturation enhancement algorithm respectively, and enhanced I array, corrected H array and enhanced S array are generated respectively;

[0020] The corrected H array, the enhanced S array and the enhanced I array are fused to generate a corrected HSI image;

[0021] converting the corrected HSI image into an image in RGB form, and generating the target profile image, the converting including:

[0022] (R,G,B) = ((R' + m) x 255, (G' + m) x 255, (B' + m) x 255),

[0023] wherein,

[0024] C = Inorm' · S'

[0025] X = C · (1 - |(H' x 6) mod 2 - 1|)

[0026]

[0027] wherein, (R, G, B) represents the RGB values of the target profile image in RGB form after conversion, R', G', B', C, X and m are intermediate calculation variables, Inorm' represents the enhanced I array, S' represents the enhanced S array, and H' represents the corrected H array.

[0028] In one specific implementation, the processing of the normalized I array, the normalized H array and the normalized S array by the adaptive contrast enhancement algorithm, the tone correction algorithm and the saturation enhancement algorithm respectively, and the generation of the enhanced I array, the corrected H array and the enhanced S array respectively, include:

[0029] dividing the image corresponding to the normalized I array into a plurality of I channel grid blocks according to a preset size;

[0030] calculating the minimum pixel value, the maximum pixel value, the grid block mean value and the grid block standard deviation value of each I channel grid block;

[0031] adjusting the normalized I array according to the minimum pixel value, the maximum pixel value, the grid block mean value and the grid block standard deviation value, and setting the adjusted I array as the enhanced I array, the adjusting including:

[0032]

[0033] wherein, Inorm' n represents the nth pixel value after adjustment, Inorm n represents the nth pixel value of the image corresponding to the normalized I array, I min represents the minimum pixel value, I max represents the maximum pixel value, and L represents the maximum range of pixel values.

[0034] calculating a local mean value and a local standard deviation value of an H channel of a preset region of the normalized H array, and a global mean value and a global standard deviation value of an H channel of the normalized H array as a whole;

[0035] correcting the normalized H array according to the local mean value and the local standard deviation value of the H channel, the global mean value and the global standard deviation value of the H channel, and setting the corrected H array as the corrected H array, and the correction manner comprises:

[0036]

[0037] wherein H′ n represents the nth value of the corrected H array, H norm，n represents the nth value of the normalized H array, μ H represents the local mean value of the H channel, σ H represents the local standard deviation value of the H channel, represents the global standard deviation value of the H channel, represents the global mean value of the H channel;

[0038] calculating a local mean value and a local standard deviation value of an S channel of a preset region of the normalized S array;

[0039] adjusting the normalized S array according to the local mean value and the local standard deviation value of the S channel, and setting the adjusted S array as the enhanced S array, and the adjustment manner comprises:

[0040]

[0041] wherein α and β represent adaptive adjustment parameters, S′ n represents the nth value of the enhanced S array, S norm，n represents the nth value of the normalized S array, μ S represents the local mean value of the S channel, σ S represents the local standard deviation value of the S channel.

[0042] In one specific implementation, the training manner of the profile enhancement neural network model comprises:

[0043] obtaining a set of clear material accumulation profile images and corresponding feeding text information corresponding to the set of clear material accumulation profile images, the set of clear material accumulation profile images being a set of clear material accumulation profile images of a material port;

[0044] performing preprocessing on each image in the set of clear material accumulation profile images by using a histogram equalization algorithm, and generating a set of equalized profile images;

[0045] adding preset times of Gaussian noise to each image in the equalized contour image set, and generating a blurred contour image set;

[0046] fusing each image in the blurred contour image set with corresponding feeding text information by using a contour enhancement neural network model, and generating a fusion image set;

[0047] de-noising each image in the fusion image set by using a Bayesian de-noising algorithm, and generating a de-noised contour image set;

[0048] calculating an image data difference between each image in the de-noised contour image set and a corresponding image in the equalized contour image set;

[0049] if the image data difference is within a preset range, the contour recognition neural network model is trained.

[0050] In one specific implementation, the fusing each image in the blurred contour image set with corresponding feeding text information by using a contour enhancement neural network model, and generating a fusion image set includes:

[0051] encoding the feeding text information, and generating feeding text encoding data;

[0052] fusing data of each image in the blurred contour image set with corresponding feeding text encoding data, and generating image-text fusion data;

[0053] converting the image-text fusion data into a fusion image by using a contour enhancement neural network model, and generating the fusion image set.

[0054] In one specific implementation, after the output clear contour image, further comprising:

[0055] obtaining feeding text information corresponding to the original contour image, fusing the clear contour image with the feeding text information corresponding to the original contour image, and generating edge prediction original data;

[0056] inputting the edge prediction original data into a trained edge prediction neural network model, and outputting a predicted edge contour image.

[0057] In one specific implementation, the training method of the edge prediction neural network model includes:

[0058] constructing an edge prediction neural network model, the edge prediction neural network model including six convolutional layers and five pooling layers;

[0059] inputting the edge prediction original data into the edge prediction neural network model for feature extraction, and outputting a predicted edge image;

[0060] According to the material type, a loss function is constructed, and the edge prediction neural network model is optimized according to the loss function, to generate a trained edge prediction neural network model.

[0061] In a specific implementation, the inputting of the edge prediction original data into the edge prediction neural network model for feature extraction and the outputting of a predicted edge image include:

[0062] The data output by the seventh convolutional layer is 4 times up-sampled to generate seventh convolutional layer up-sampled data, and the data output by the fourth pooling layer is 2 times up-sampled to generate fourth pooling layer up-sampled data;

[0063] The seventh convolutional layer up-sampled data, the fourth pooling layer up-sampled data and the data output by the third pooling layer are fused to generate convolutional pooling fusion data;

[0064] The convolutional pooling fusion data is 8 times up-sampled to output the predicted edge image.

[0065] In a second aspect, the present application provides a powder pneumatic conveying image processing device, which adopts the following technical scheme: the device comprises:

[0066] An image acquisition module is configured to acquire an image of a material accumulation profile at a feeding port and corresponding feeding text information, and set the image of the material accumulation profile at the feeding port as an original profile image;

[0067] An equalization processing module is configured to pre-process the original profile image by using a histogram equalization algorithm and generate a target profile image;

[0068] A noise adding module is configured to add Gaussian noise of a preset number of times to the target profile image to generate a noise profile image;

[0069] An image output module is configured to input the noise profile image and the corresponding feeding text information into a trained profile enhancement neural network model, and output a clear profile image after Bayesian denoising.

[0070] In a third aspect, the present application provides a computer device, which adopts the following technical scheme: comprising a memory and a processor, the memory stores a computer program capable of being loaded and executed by the processor to perform any of the above powder pneumatic conveying image processing methods.

[0071] In a fourth aspect, the present application provides a computer readable storage medium, which adopts the following technical scheme: storing a computer program capable of being loaded and executed by a processor to perform any of the above powder pneumatic conveying image processing methods.

[0072] In summary, the present application has the following beneficial technical effects:

[0073] The original contour image is pre-processed and then input into the trained neural network model for calculation, and a clear contour image is output. The clear contour image is an image with clear contours and clear edges. Based on the clear contour image, the stacking trend of the discharge port can be determined, thereby avoiding blockage as much as possible, and improving the feeding efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0074] Figure 1 is a flowchart of the powder pneumatic conveying image processing method in the embodiment of the present application;

[0075] Figure 2 is a framework schematic diagram of the contour enhancement neural network model training in the embodiment of the present application;

[0076] Figure 3 is a convolutional network structure schematic diagram of the contour enhancement neural network model in the embodiment of the present application;

[0077] Figure 4 is a feature extraction schematic diagram of the edge prediction neural network model in the embodiment of the present application;

[0078] Figure 5 is a schematic diagram of the powder pneumatic conveying image processing device in the embodiment of the present application;

[0079] Figure 6 is a schematic diagram of the computer device in the embodiment of the present application.

[0080] Reference signs: 501, image acquisition module; 502, equalization processing module; 503, noise adding module; 504, image output module. DETAILED DESCRIPTION

[0081] The following will be described in detail in combination with Figures 1-6 The present application will be further described in detail.

[0082] The embodiment of the present application discloses a powder pneumatic conveying image processing method. The method is applied to a powder pneumatic conveying image processing system. The image processing method can process the image of the material stacking contour of the feeding port in the powder pneumatic conveying feeding process, so that the contour of the image is clearer and the contour edge is clearer. In this way, the blockage problem in the mechanical arm feeding process can be avoided as much as possible, and the efficiency of the mechanical arm feeding is improved.

[0083] In the pharmaceutical, food, chemical, new energy and other industries, ton bag, small bag of powder or granular material in the feeding process, usually artificial feeding way, or using mechanical arm or other conveyors to feeding port feeding. Artificial feeding way is generally low efficiency, mechanical arm or other conveyors to feeding port feeding can solve the problem of low efficiency of artificial feeding, but mechanical arm feeding, may be difficult to find timely blockage problem, thus may be more impact efficiency, at the same time cause material waste. There are many conventional parameters to monitor the working state of the feeding port, including feeding port level meter, flow, pipe gas pressure value, etc., which are usually monitored by video image; however, the video image of the feeding port working state is blurred due to the dust generated by the powder during work, which causes the video image of the camera to be blurred.

[0084] At present, there are some image data processing methods to process the image of the feeding port to obtain clear video image; however, after processing by the current image processing method, it is still difficult to see the material accumulation profile of the feeding port clearly through the image, and it is also difficult to see the dynamic change process of the material accumulation profile of the feeding port, so as to cause difficulty in judging whether the discharge port has a serious tendency of accumulation based on the material accumulation profile of the feeding port, thereby causing blockage, affecting the feeding efficiency, and also possibly affecting the material pollution, causing economic loss. In order to help improve the clarity of the profile image, the dynamic change process of the accumulation profile can be seen from the image, and the feeding efficiency is improved, the application provides a powder pneumatic conveying image processing method.

[0085] Reference Figure 1 The method comprises the following steps:

[0086] S10, obtaining the image of the material accumulation profile of the feeding port and the corresponding feeding text information, and setting the image of the material accumulation profile of the feeding port as the original profile image.

[0087] Specifically, generally, the user will monitor the working state of the feeding port by the way of camera shooting video image, the video image shot by the camera is processed to obtain each frame of image, the obtained frame image of the video image is input into the image processing system, the system obtains the image of the material accumulation profile of the feeding port, and sets the image of the material accumulation profile of the feeding port as the original profile image. The corresponding feeding text information refers to the type of material corresponding to the image, the temperature, height, air pressure, humidity, light condition, particle matter (PM2.5) value and other external factors in the current situation of the image are obtained, for example, the text information obtained by the system can be expressed as "flour 36℃ 57cm 1009hpa 36% 20% 122μg / m 3text information and the contour image together into the neural network model can increase the weight of the text information, so that the contour image can be influenced to obtain a contour image with higher clarity.

[0088] S20, an adaptive contrast enhancement fusion color correction algorithm is used to preprocess the original contour image and generate a target contour image.

[0089] Specifically, the adaptive contrast enhancement fusion color correction algorithm is used to preprocess the original contour image, and the adaptive contrast enhancement fusion color correction algorithm includes an adaptive contrast enhancement algorithm and a color correction fusion algorithm, so that the brightness and color distribution of the image can be adjusted more intelligently.

[0090] S30, adding a preset number of Gaussian noises to the target contour image to generate a noise contour image.

[0091] Specifically, a preset number of Gaussian noises are added to the target image obtained after preprocessing, wherein the Gaussian noise is added in the form of cumulative noise, which can also be understood as adding Gaussian noise to the target image for the first time to obtain an image with once-added Gaussian noise, adding Gaussian noise to the image with once-added Gaussian noise for the second time to obtain an image with twice-added Gaussian noise, and so on, to obtain an image with a preset number of added Gaussian noises. Specifically, for example, assuming that the target image obtained after preprocessing is X0, the number of added Gaussian noises is k, and generally, the range of k can be set to [5, 15], then the noise image obtained after adding the t-th Gaussian noise can be represented as:

[0092]

[0093] wherein X0 represents the target image before adding noise; X t represents the noise image after adding the t-th Gaussian noise; Z t represents the mean of the Gaussian random variable added from 1 to t, and all the added Gaussian random variables follow the Gaussian distribution N(0, 1), and the mean Z t also follows the Gaussian distribution N(0, 1); a t represents the weight size of the Gaussian noise in the noise image obtained by adding the (t-1)th Gaussian noise, a t The calculation method of a

[0094] a t = 1-β t

[0095] wherein β t is a pre-set value, and the range of β t is [0.0001, 0.02].t With the increase of t, increase in proportion, the initial value is 0.0001, and increase to the maximum value 0.02; after k times of superposition, the final noise profile image is obtained.

[0096] S40, input the noise profile image and the corresponding feeding text information into the trained contour enhancement neural network model, and output the clear contour image after Bayesian denoising.

[0097] Specifically, the obtained noise profile image and the corresponding feeding text information obtained are input into the contour enhancement neural network model which has been trained, and the trained neural network model is used to increase the weight influence of the text information according to the influence of the text information on the contour image information, which is helpful to obtain a contour image with higher clarity. Since the input is a noise contour image, the image output by the model needs to be denoised to obtain a clear contour image. Specifically, the denoising can use Bayesian denoising. For example, suppose the image output by the neural network with noise is Y k , and the image obtained after k times of noise removal is Y0, where k times of denoising corresponds to k times of adding noise; the image Y t-1 obtained after the (t-1)th denoising can be represented as:

[0098]

[0099] Where a t represents the weight size of the noise in the noise image obtained after the (t-1)th denoising, which is consistent with the meaning of a t in the added Gaussian noise; θ t represents a variable parameter, the initial value is 1, and can be adjusted according to the difference between the training result and the actual target; σ t represents an adjustment coefficient, and the calculation method of σ t can be represented as:

[0100]

[0101] Finally, the clear contour image is obtained after k times of denoising.

[0102] In the scheme of the present application, the original contour image is first preprocessed before being input into the trained neural network model for calculation, and a clear contour image is output, and the clear contour image obtained is an image with clear contour and clear edge, and based on the clear contour image, the stacking trend of the discharge port can be judged, so as to avoid material blocking as much as possible, and thus improve the feeding efficiency. The preprocessing adopts a histogram equalization algorithm, which is a preprocessing method for dust in a relatively dark environment, and can improve the image quality, so that the image after preprocessing can show more details hidden due to the external environment.

[0103] In one embodiment, the original contour image is a color image, the original contour image is in RGB form, the adaptive contrast enhancement fusion color correction algorithm is used to preprocess the original contour image, and the step of generating the target contour image can be specifically implemented as:

[0104] Firstly, the original contour image in RGB form is converted into an original contour image in HSI form. The original contour image in HSI form includes H channel, S channel and I channel, which respectively represent hue (Hue), saturation (Saturation) and intensity (Intensity), and the conversion calculation formula is as follows:

[0105]

[0106] H = 360°-H (if B > G)

[0107]

[0108] Wherein, R, G and B respectively represent the data of R channel, G channel and B channel of the original contour image in RGB form, H, S and I respectively represent the data of H channel, S channel and I channel of the converted original contour image in HSI form;

[0109] Subsequently, the data of H channel, S channel and I channel are normalized respectively to generate normalized H array, normalized S array and normalized I array, and the calculation formula of normalization can be expressed as:

[0110] S norm = S,

[0111] Wherein, H norm represents the normalized H array, S norm represents the normalized S array, I norm represents the normalized I array, and I max represents the maximum value in the I channel data;

[0112] After that, the adaptive contrast enhancement algorithm, the hue correction algorithm and the saturation enhancement algorithm are used to process the normalized I array, the normalized H array and the normalized S array respectively, and the enhanced I array, the corrected H array and the enhanced S array are generated; specifically, the normalized I array I norm The adaptive contrast enhancement algorithm is applied to the normalized H array and the normalized S array, and the color correction fusion algorithm is used to process them.

[0113] Firstly, the image corresponding to the normalized I array is divided into several I channel grid blocks according to the preset size, and the size of each region can be adjusted according to the image resolution and characteristics. After that, local statistical calculation is performed to calculate the minimum pixel value, the maximum pixel value, the grid block mean value and the grid block standard deviation value of each I channel grid block. Finally, adaptive adjustment is performed to adjust the normalized I array according to the minimum pixel value, the maximum pixel value, the grid block mean value and the grid block standard deviation value, and the adjusted I array is set as the enhanced I array, and the adjustment method includes:

[0114]

[0115] wherein Inorm' n represents the n th pixel value of the normalized I array, Inorm n represents the n th pixel value of the image corresponding to the normalized I array, I min represents the minimum pixel value, I max represents the maximum pixel value, and L represents the maximum range of pixel value.

[0116] After that, the data of the H channel is processed. Firstly, local statistical calculation is performed to calculate the H channel local mean value and the H channel local standard deviation value of the preset region of the normalized H array, and the H channel global mean value and the H channel global standard deviation value of the normalized H array globally. After that, hue equalization processing is performed to correct the normalized H array according to the H channel local mean value, the H channel local standard deviation value, the H channel global mean value and the H channel global standard deviation value, and the corrected H array is set as the corrected H array, and the correction method includes:

[0117]

[0118] wherein H' n represents the n th value of the corrected H array, H norm,n represents the n th value of the normalized H array, μ H represents the H channel local mean value, σ H represents the H channel local standard deviation value, represents the H channel global standard deviation value, represents the H channel global mean value; the H channel global mean value and the H channel global standard deviation value are used to maintain the consistency of the overall hue.

[0119] After that, the data of the S channel is processed. First, local statistical calculation is performed to calculate the S channel local mean value and the S channel local standard deviation value of the preset region of the normalized S array. Then, saturation enhancement processing is performed, and the normalized S array is adjusted according to the S channel local mean value and the S channel local standard deviation value. The adjusted S array is set as the enhanced S array, and the adjustment method includes:

[0120]

[0121] wherein a and b represent adaptive adjustment parameters, which are dynamically adjusted according to the saturation distribution of the local region to ensure that the saturation is enhanced within a reasonable range, S' n represents the nth value of the enhanced S array, S norm,n represents the nth value of the normalized S array, m S represents the S channel local mean value, s S represents the S channel local standard deviation value.

[0122] After processing the data of each channel, the corrected H array, the enhanced S array and the enhanced I array are fused to generate a corrected HSI image. The corrected HSI image can be represented as:

[0123]

[0124] wherein HSI new represents the corrected HSI image, H' represents the corrected H array, S' represents the enhanced S array, and Inorm' represents the enhanced I array.

[0125] Finally, the corrected HSI image is converted into an image in RGB form, and a target contour image is generated. The conversion method includes:

[0126] (R, G, B) = ((R' + m) x 255, (G' + m) x 255, (B' + m) x 255),

[0127] wherein

[0128] C = Inorm' x S'

[0129] X = C x (1 - |(H' x 6) mod 2 - 1|)

[0130]

[0131] wherein (R, G, B) represents the RGB value of the converted target contour image in RGB form, R', G', B', C, X and m are intermediate calculation variables, Inorm' represents the enhanced I array, S' represents the enhanced S array, and H' represents the corrected H array.

[0132] The adaptive contrast enhancement fusion color correction algorithm is used for preprocessing the original image, which can more intelligently adjust the brightness and color distribution of the image, mainly including: 1) local adaptive enhancement: according to the brightness and color distribution of the local area of the image, the contrast and color are adaptively adjusted to avoid over-enhancement or detail loss caused by global adjustment; 2) tone balance and saturation enhancement: through tone balance, the color distribution is more uniform, and saturation enhancement enhances the vividness of image color, while avoiding color distortion; 3) maintain natural color: by introducing global and local statistical parameters, the image after color adjustment is ensured to maintain natural and realistic feeling.

[0133] The traditional preprocessing method is to directly perform histogram equalization processing on the image in RGB form. However, the obtained image of the material accumulation profile of the feeding port has the problems of single target, more details, and insufficient light, so the traditional histogram containing RGB color channel is not needed for image processing. Based on the sensitivity of light, after converting the RGB image into an HSI image, the adaptive contrast enhancement fusion color correction algorithm is used for processing, which can more effectively optimize the brightness and color distribution, obtain a better histogram, and thus improve the image quality and display more details hidden due to the external environment.

[0134] In one embodiment, with reference to Figure 2 , the step of training the contour enhancement neural network model can be specifically implemented as:

[0135] First, a clear material accumulation profile image set and corresponding feeding text information corresponding to the clear material accumulation profile image set are obtained, and the clear material accumulation profile image set is a set of clear material accumulation profile images of the feeding port. The image set input by the model is a set of frame images split from the video image, and the corresponding feeding text information refers to the type of material corresponding to the image. The temperature, height, air pressure, humidity, light conditions, and particle matter (PM2.5) values of the current situation of the image are obtained. It needs to be clear that the image input by the training model is a clear image of the material accumulation profile of the feeding port, and the size of all input images is 512*512. The input text information is the existence of the correction condition for model training.

[0136] After that, the histogram equalization algorithm is used to pre-process each image in the set of clear material accumulation profile images, and a set of equalized profile images is generated. The pre-processing method of obtaining the set of equalized profile images is consistent with the pre-processing method of obtaining the target profile image. First, the image in RGB form is converted into an image in HSI form. After normalizing the data in the H, S, and I channels of the image in HSI form, the histogram equalization processing algorithm is used. Then, the set of images in HSI form is converted into a set of images in RGB form. Finally, the set of equalized profile images is obtained. Since the powder pneumatic conveying environment is relatively closed, pre-processing the image can significantly reduce the training difficulty. Each image in the set of equalized profile images has a size of 512*512 after pre-processing. After pre-processing, a preset number of Gaussian noise is added to each image in the set of equalized profile images, and a set of fuzzy profile images is generated. Specifically, the way of adding Gaussian noise is consistent with the way of obtaining the noise profile image in step S30. After adding noise several times, the fuzzy profile image is obtained. The size of the processed image is 512*512*3, where 3 represents the R, G, and B channels. The set of fuzzy profile images is the fuzzy profile image set.

[0137] After adding Gaussian noise, the profile enhancement neural network model is used to fuse each image in the set of fuzzy profile images with the corresponding feeding text information, and a set of fusion images is generated. Specifically, the feeding text information is first encoded, and the feeding text encoded data is generated. The text information obtained by the system is in the format of "flour 36℃ 57cm 1009hpa 36% 20% 122μg / m 3 ". Each parameter is separated by a space. For text input data, no further segmentation processing is required, which reduces the workload and increases the feature extraction efficiency. Since the input text information contains Chinese information, the characters need to be encoded first. Utf-8 is used for encoding. Then, a 3*1 convolution kernel is used to adjust the size of the text information. The final size of the text information is 512*1. Then, the data of each image in the set of fuzzy profile images is fused with the corresponding feeding text encoded data to generate image text fusion data. The size of the fusion data is 512*512*4, where 4 represents the R, G, B, and text data channels. Finally, the profile enhancement neural network model is used to convert the image text fusion data into fusion images, and a set of fusion images is generated. The purpose of using neural network fusion data is to increase the weight of text information, so that the weight can affect the image information. Specifically, refer to Figure 3The convolution kernel network can be designed as an eight-layer convolution structure, which is divided into four layers of down-sampling and four layers of up-sampling. A nonlinear transformation is added between each convolution layer to prevent feature redundancy between convolution layers. Second, for example, data transmission of a 256*256*32 layer to a 256*256*16 layer is used to transmit the features extracted in the previous step to the subsequent convolution layer as a reference to improve the accuracy of the model. After the output of the neural network model, the size of each image obtained is 512*512*3.

[0138] After obtaining the fused image set, each image in the fused image set is denoised using the Bayesian denoising algorithm to generate a denoised contour image set. The specific Bayesian denoising method is consistent with the Bayesian denoising method used in step S40. After adding several times of Gaussian noise, the corresponding number of denoising is performed, and the specific denoising method is not described here. Finally, the image data difference between each image in the denoised contour image set and the corresponding image in the equalized contour image set is calculated. For example, assuming that the denoised contour image set is Y, the equalized contour image set is X, and the number of images in the image set is N, the image data difference calculation method can be represented as:

[0139]

[0140] Where i represents the i-th image; if the image data difference is within the preset range, the contour recognition neural network model training is completed, and the preset value is the value of loss1. In the present embodiment, the preset value is set to 20, and the size of the preset value can be set according to the actual image processing situation, which is not limited here.

[0141] It should be noted that in the present embodiment, an image set is taken as an example for illustration, and an image set can be understood as an image set obtained from a video image. In actual application, multiple video images can be used for model training, which can increase the data samples and improve the accuracy of the trained model.

[0142] In the present application, the contour enhancement neural network model with high precision is obtained by training and image pair model training. In actual application, the trained neural network model can obtain clear edge images without artifacts.

[0143] In one embodiment, considering that the material descends at a relatively fast speed, the edge changes have strong real-time performance, and the image processing neural network occupies a certain amount of calculation time and calculation resources, which may result in a long time consumption. Therefore, after outputting the clear contour image, the following steps can be performed:

[0144] The text information corresponding to the original contour image is obtained, the text information corresponding to the clear contour image and the original contour image is fused, and edge prediction original data is generated. Specifically, the text information includes the type of the corresponding material in the image, the temperature, the height, the air pressure, the humidity, the light condition, and the value of particulate matter (PM2.5) under the current condition of the image, and the like. For example, the text information obtained by the system can be represented as “flour 36℃ 57cm 1009hpa 36% 20% 122μg / m 3 ”, since the text information contains Chinese, the text information also needs to be encoded first. The size of the obtained clear contour image is 512*512*3, and the size of the edge prediction original data obtained after combining with the text data is 512*512*4; then, the edge prediction original data is input into the trained edge prediction neural network model, and a predicted edge contour image is output.

[0145] In the scheme of the present application, since the material descends at a relatively high speed, the change of the edge has a strong real-time performance, and the neural network for image processing occupies a certain amount of calculation time and calculation resources, which may cause a long time consumption. In order to achieve effective real-time monitoring, an edge prediction neural network model with a short time consumption and a high accuracy is designed and trained to predict the edge contour; in addition, the addition of text data can greatly reduce the data required for reasoning.

[0146] In one embodiment, the training of the edge prediction neural network model can be specifically performed as follows:

[0147] First, an edge prediction neural network model is constructed, which includes six convolutional layers and five pooling layers. The connection order of the convolutional layers and the pooling layers in the constructed edge prediction neural network model can be: a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, a fourth convolutional layer, a fourth pooling layer, a fifth convolutional layer, a fifth pooling layer, a sixth convolutional layer, and a seventh convolutional layer. The size of the input data is 512*512*4. The input data is subjected to a convolution operation to obtain a first convolutional layer, and the size of the first convolutional layer is 512*512*4. Then, a pooling operation is performed to obtain data with a size of 256*256*16. Subsequently, a convolution operation is performed to obtain a second convolutional layer with a size of 256*256*16. After pooling, the size becomes 128*128*64. The convolution and pooling operations are continuously performed, and the data required to be input into the prediction network is 64*64*256. The above operations are continuously performed, and 32*32*1024 is obtained. In this way, the fifth pooling layer becomes 16*16*4096. Subsequently, two convolution operations are performed, and the data obtained after the two convolution operations is 16*16*4096.

[0148] Then, the edge prediction original data is input into the edge prediction neural network model for feature extraction, and a predicted edge image is output. Specifically, referring to Figure 4 The data output by the seventh convolutional layer is 4 times up-sampled to generate seventh convolutional layer up-sampled data, and the data output by the fourth pooling layer is 2 times up-sampled to generate fourth pooling layer up-sampled data. The seventh convolutional layer up-sampled data, the fourth pooling layer up-sampled data, and the data output by the third pooling layer are fused to generate convolutional pooling fusion data. The convolutional pooling fusion data is 8 times up-sampled to output the predicted edge image.

[0149] Finally, a loss function is constructed according to the material types, and the edge prediction neural network model is optimized according to the loss function to generate a trained edge prediction neural network model. The constructed loss function can be expressed as:

[0150]

[0151] Wherein, N represents (the number of input material types + 1), for example, assuming that there is only a single material in the input material this time, the value of N is 2; y i represents the label of the material type i, wherein the positive class is 1, which can be understood as this type, and the negative class is 0, which can be understood as not this type; represents the predicted value; in most cases, the material type of the negative class will not be much larger than the material type of the positive class, so the loss function method described above is used in the embodiment of the present application.

[0152] It should be noted that in the embodiment of the present application, the clear contour image set in the edge prediction original data set can directly use the clear contour image set obtained by the contour enhancement neural network model described above. The model is trained and optimized by multiple image data samples in the data set to improve the prediction accuracy of the model.

[0153] In the present application, the edge prediction neural network model with high precision is obtained by training and image pairs. In actual application, the trained neural network model can obtain an edge image with high prediction accuracy.

[0154] Figure 1 is a flowchart of a powder pneumatic conveying image processing method in one embodiment. It should be understood that although Figure 1 the steps in the flowchart of the present application are displayed in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows; unless otherwise specified in this document, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other order; and Figure 1At least one of the steps in the above method can include a plurality of sub-steps or a plurality of stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of the sub-steps or stages is not necessarily sequential, but can be executed alternately or alternately with at least one part of other steps or sub-steps or stages of other steps.

[0155] Based on the above method, the embodiment of the application further discloses a powder pneumatic conveying image processing device.

[0156] With reference to Figure 5 The device comprises the following modules:

[0157] The image acquisition module 501 is configured to acquire an image of a material accumulation profile at a feeding port and corresponding feeding text information, and set the image of the material accumulation profile at the feeding port as an original profile image.

[0158] The equalization processing module 502 is configured to pre-process the original profile image by using an adaptive contrast enhancement fusion color correction algorithm, and generate a target profile image.

[0159] The noise adding module 503 is configured to add a preset number of Gaussian noises to the target profile image, and generate a noise profile image.

[0160] The image output module 504 is configured to input the noise profile image and the corresponding feeding text information into a trained profile enhancement neural network model, and output a clear profile image after Bayesian denoising.

[0161] In one embodiment, the equalization processing module 502 is specifically configured to convert the original profile image in RGB form into an original profile image in HSI form, and the original profile image in HSI form includes an H channel, an S channel and an I channel, and the conversion manner includes:

[0162]

[0163] H = 360°-H (if B > G)

[0164]

[0165] wherein R, G and B respectively represent data of an R channel, a G channel and a B channel of the original profile image in RGB form, and H, S and I respectively represent data of an H channel, an S channel and an I channel of the original profile image in HSI form after conversion;

[0166] The data of the H channel, the S channel and the I channel are respectively normalized to generate a normalized H array, a normalized S array and a normalized I array, and the normalization calculation manner of the data of the H channel, the S channel and the I channel includes:

[0167] S norm =S,

[0168] Among them, H norm Represents the normalized H array, S norm Represents the normalized S array, I norm Represents the normalized array I, I max This represents the maximum value in the I channel data;

[0169] An adaptive contrast enhancement algorithm, a tone correction algorithm, and a saturation enhancement algorithm are used to process the normalized I array, normalized H array, and normalized S array, respectively, to generate enhanced I array, corrected H array, and enhanced S array, respectively. The corrected H array, enhanced S array, and enhanced I array are then fused to generate a corrected HSI image. The corrected HSI image is then converted to RGB format to generate a target contour image. The conversion methods include:

[0170] (R, G, B) = ((R′+m)×255, (G′+m)×255, (B′+m)×255),

[0171] in,

[0172] C = Inorm′·S′

[0173] X = C·(1-|(H′×6)mod2-1|)

[0174]

[0175] Where (R,G,B) represents the RGB values ​​of the target contour image in RGB form after conversion, R', G', B', C, X and m are intermediate calculation variables, Inorm′ represents the enhancement I array, S' represents the enhancement S array, and H' represents the correction H array.

[0176] In one embodiment, the equalization processing module 502 is specifically used to divide the image corresponding to the normalized I-array into several I-channel grid blocks of a preset size; calculate the minimum pixel value, maximum pixel value, grid block mean, and grid block standard deviation of each I-channel grid block; adjust the normalized I-array based on the minimum pixel value, maximum pixel value, grid block mean, and grid block standard deviation, and set the adjusted I-array as the enhanced I-array. The adjustment method includes:

[0177]

[0178] Among them, Inorm′ n Inorm represents the adjusted value of the nth pixel.n denotes the n-th pixel value of the normalized I array, I min denotes the minimum pixel value, I max denotes the maximum pixel value, L denotes the maximum range of pixel values;

[0179] calculating the H channel local mean and the H channel local standard deviation of the preset region of the normalized H array, and the H channel global mean and the H channel global standard deviation of the normalized H array globally; correcting the normalized H array according to the H channel local mean, the H channel local standard deviation, the H channel global mean and the H channel global standard deviation, and setting the corrected H array as a corrected H array, and the correction mode includes:

[0180]

[0181] wherein H' = H - H n denotes the n-th value of the corrected H array, H norm，n denotes the n-th value of the normalized H array, μ H denotes the H channel local mean, σ H denotes the H channel local standard deviation, denotes the H channel global standard deviation, denotes the H channel global mean;

[0182] calculating the S channel local mean and the S channel local standard deviation of the preset region of the normalized S array; adjusting the normalized S array according to the S channel local mean and the S channel local standard deviation, and setting the adjusted S array as an enhanced S array, and the adjustment mode includes:

[0183]

[0184] wherein α and β denote adaptive adjustment parameters, S' = S - S n denotes the n-th value of the enhanced S array, S norm,n denotes the n-th value of the normalized S array, μ S denotes the S channel local mean, σ S denotes the S channel local standard deviation.

[0185] In one embodiment, the image output module 504 is further configured to obtain a set of clear material accumulation profile images and corresponding feeding text information corresponding to the set of clear material accumulation profile images, the set of clear material accumulation profile images being a collection of material hopper material accumulation profile images; pre-process each image in the set of clear material accumulation profile images using a histogram equalization algorithm, and generate a set of equalized profile images; add a preset number of Gaussian noises to each image in the set of equalized profile images, and generate a set of blurred profile images; fuse each image in the set of blurred profile images with the corresponding feeding text information using a profile enhancement neural network model, and generate a set of fused images; denoise each image in the set of fused images using a Bayesian denoising algorithm, and generate a set of denoised profile images; calculate an image data difference between each image in the set of denoised profile images and the corresponding image in the set of equalized profile images; and if the image data difference is within a preset range, the profile recognition neural network model is trained.

[0186] In one embodiment, the image output module 504 is further configured to encode the feeding text information, and generate feeding text encoded data; fuse the data of each image in the set of blurred profile images with the corresponding feeding text encoded data, and generate image-text fusion data; and convert the image-text fusion data into fused images using the profile enhancement neural network model, and generate a set of fused images.

[0187] In one embodiment, the image output module 504 is further configured to obtain feeding text information corresponding to the original profile image, fuse the clear profile image with the feeding text information corresponding to the original profile image to generate edge prediction original data; input the edge prediction original data into the trained edge prediction neural network model, and output a predicted edge profile image.

[0188] In one embodiment, the image output module 504 is further configured to construct an edge prediction neural network model, the edge prediction neural network model including six convolutional layers and five pooling layers; input the edge prediction original data into the edge prediction neural network model for feature extraction, and output a predicted edge image; construct a loss function according to the material type, and optimize the edge prediction neural network model according to the loss function to generate a trained edge prediction neural network model.

[0189] In one embodiment, the image output module 504 is further configured to perform 4 times up-sampling on the data output by the seventh convolutional layer to generate seventh convolutional layer up-sampling data, and perform 2 times up-sampling on the data output by the fourth pooling layer to generate fourth pooling layer up-sampling data; fuse the seventh convolutional layer up-sampling data, the fourth pooling layer up-sampling data, and the data output by the third pooling layer to generate convolutional-pooling fusion data; and perform 8 times up-sampling on the convolutional-pooling fusion data to output a predicted edge image.

[0190] The powder pneumatic conveying image processing device provided in the embodiments of the present application can be applied to the powder pneumatic conveying image processing method provided in the above embodiments, and the related details refer to the above method embodiments, which have similar implementation principles and technical effects, and thus will not be described here again.

[0191] It should be noted that the powder pneumatic conveying image processing device provided in the embodiments of the present application is only exemplified by the above division of the functional modules / functional units when performing powder pneumatic conveying image processing, and in actual application, the above functions can be completed by different functional modules / functional units according to needs, that is, the internal structure of the powder pneumatic conveying image processing device is divided into different functional modules / functional units to complete all or part of the above described functions. In addition, the implementation manner of the powder pneumatic conveying image processing method provided in the above method embodiments and the implementation manner of the powder pneumatic conveying image processing device provided in the present embodiment belong to the same concept, and the specific implementation process of the powder pneumatic conveying image processing device provided in the present embodiment is described in the above method embodiments, which will not be described here again.

[0192] The embodiments of the present application also disclose a computer device.

[0193] Specifically, as shown in the figure, Figure 6 The computer device can be a desktop computer, a notebook computer, a palm computer, a cloud server, and the like. The computer device can include, but is not limited to, a processor and a memory. The processor and the memory can be connected through a bus or other manners. The processor can be a central processing unit (CPU). The processor can also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, graphics processing units (GPUs), embedded neural-network processing units (NPUs) or other dedicated deep learning co-processors, discrete gates or transistor logic devices, discrete hardware components, and the like chips, or combinations of the above chips.

[0194] The memory, as a non-transitory computer readable storage medium, can be used to store non-transitory software programs, non-transitory computer executable programs and modules, such as program instructions / modules corresponding to the methods in the above embodiments of the present application. The processor executes various functions and data processing of the processor by running the non-transitory software programs, instructions and modules stored in the memory, that is, implements the methods in the above method embodiments. The memory can include a program storage area and a data storage area, wherein the program storage area can store an operating system and application programs required by at least one function; and the data storage area can store data created by the processor and the like. In addition, the memory can include a high-speed random access memory, and can also include a non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state memory device. In some embodiments, the memory can optionally include a memory remotely arranged with respect to the processor, and these remote memories can be connected to the processor through a network. Examples of the above network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0195] The embodiments of the present application also disclose a computer readable storage medium.

[0196] Specifically, the computer readable storage medium is used to store a computer program, and the computer program is executed by the processor to implement the methods in the above method embodiments. Those skilled in the art can understand that all or part of the processes in the above method embodiments of the present application can be completed by a computer program instructing related hardware, and the program can be stored in a computer readable storage medium. When the program is executed, it can include the processes of the above method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), a flash memory, a hard disk drive (HDD) or a solid-state drive (SSD), etc. The storage medium can also include a combination of the above types of memories.

[0197] The specific embodiments are only an explanation of the present application, and are not a limitation of the present application. Those skilled in the art can make modifications to the embodiments without creative contributions as long as the modifications are within the scope of the claims of the present application.

Claims

1. A powder pneumatic conveying image processing method, characterized by: The method comprises: acquire the image of the material accumulation profile of the feeding port and the corresponding feeding text information, and set the image of the material accumulation profile of the feeding port as an original profile image; use an adaptive contrast enhancement fusion color correction algorithm to preprocess the original profile image and generate a target profile image; add a preset number of Gaussian noises to the target profile image to generate a noise profile image; input the noise profile image and the corresponding feeding text information into a trained profile enhancement neural network model, and output a clear profile image after Bayesian denoising; the original profile image is a color image, the original profile image is in RGB form, and the adaptive contrast enhancement fusion color correction algorithm is used to preprocess the original profile image and generate a target profile image, which comprises: convert the original profile image in RGB form into an original profile image in HSI form, the original profile image in HSI form comprises an H channel, an S channel and an I channel, and the conversion mode comprises: ， wherein R, G and B respectively represent the data of the R channel, the G channel and the B channel of the original profile image in RGB form, and H, S and I respectively represent the data of the H channel, the S channel and the I channel of the converted original profile image in HSI form; normalize the data of the H channel, the S channel and the I channel respectively to generate a normalized H array, a normalized S array and a normalized I array, and the normalization calculation mode of the data of the H channel, the S channel and the I channel comprises: ， where H norm represents the normalized H array, S norm represents the normalized S array, I norm represents the normalized I array, I max represents the maximum value in the I channel data; use an adaptive contrast enhancement algorithm, a hue correction algorithm and a saturation enhancement algorithm to process the normalized I array, the normalized H array and the normalized S array respectively, and generate an enhanced I array, a corrected H array and an enhanced S array respectively; fuse the corrected H array, the enhanced S array and the enhanced I array to generate a corrected HSI image; convert the corrected HSI image into an image in RGB form to generate the target profile image, and the conversion mode comprises: ， wherein , ， where (R, G, B) represents the RGB values of the target profile image in converted RGB form, R', G', B', C, X, and m are intermediate calculation variables, represents the enhanced I array, S' represents the enhanced S array, and H' represents the corrected H array. use an adaptive contrast enhancement algorithm, a hue correction algorithm and a saturation enhancement algorithm to process the normalized I array, the normalized H array and the normalized S array respectively, and generate an enhanced I array, a corrected H array and an enhanced S array respectively, which comprises: divide the image corresponding to the normalized I array into a plurality of I channel grid blocks according to a preset size; calculate the minimum pixel value, the maximum pixel value, the grid block mean value and the grid block standard deviation value of each I channel grid block; adjust the normalized I array according to the minimum pixel value, the maximum pixel value, the grid block mean value and the grid block standard deviation value, and set the adjusted I array as the enhanced I array, and the adjustment mode comprises: ， wherein, denotes the adjusted nth pixel value, denotes the nth pixel value of the image corresponding to the normalized I array, I min denotes the minimum pixel value, I max denotes the maximum pixel value, L denotes the maximum range of pixel values; calculate the H channel local mean value and the H channel local standard deviation value of the preset region of the normalized H array, and the H channel global mean value and the H channel global standard deviation value of the normalized H array as a whole; The normalized H array is corrected based on the local mean, local standard deviation, global mean, and global standard deviation of the H channels, and the corrected H array is set as the corrected H array. The correction method includes: ， wherein, denotes the n-th value of the corrected H array, denotes the n-th value of the normalized H array, denotes the local mean value of the H channel, denotes the local standard deviation value of the H channel, denotes the global standard deviation value of the H channel, denotes the global mean value of the H channel; Calculate the local mean and local standard deviation of the S-channel in the preset region of the normalized S-array; The normalized S-array is adjusted based on the local mean and local standard deviation of the S-channel, and the adjusted S-array is set as the enhanced S-array. The adjustment method includes: ， wherein a and β represent adaptive adjustment parameters, denotes the n-th value of the enhanced S array, denotes the n-th value of the normalized S array, denotes the local mean value of the S channel, denotes the local standard deviation value of the S channel.

2. The method of claim 1, wherein: The training methods for the contour enhancement neural network model include: Obtain a set of images with clear material accumulation outlines and corresponding feeding text information, wherein the set of images with clear material accumulation outlines is a collection of images with clear material accumulation outlines at the feed inlet. The histogram equalization algorithm is used to preprocess each image in the set of images with clear material accumulation contours, and an equalized contour image set is generated. Add a preset number of Gaussian noises to each image in the equalized contour image set to generate a blurred contour image set; A contour enhancement neural network model is used to fuse each image in the fuzzy contour image set with the corresponding material feeding text information, and a fused image set is generated. The Bayesian denoising algorithm is used to denoise each image in the fused image set to generate a denoised contour image set; Calculate the image data difference between each image in the denoised contour image set and the corresponding image in the equalized contour image set; If the difference in the image data is within a preset range, then the contour enhancement neural network model training is complete.

3. The method of claim 2, wherein: The step of using a contour enhancement neural network model to fuse each image in the blurred contour image set with the corresponding material feeding text information and generate a fused image set includes: The material feeding text information is encoded, and material feeding text encoding data is generated; The data of each image in the blurred contour image set is fused with the corresponding material feeding text encoding data to generate image text fusion data; The image-text fusion data is converted into fused images using a contour enhancement neural network model, and the fused image set is generated.

4. The method of claim 1, wherein: Following the output of the clear outline image, the following is also included: Obtain the material feeding text information corresponding to the original contour image, and fuse the clear contour image with the material feeding text information corresponding to the original contour image to generate edge prediction raw data; The raw edge prediction data is input into the trained edge prediction neural network model, and the predicted edge contour image is output.

5. The method of claim 4, wherein: The training methods for the edge prediction neural network model include: Construct an edge prediction neural network model, which includes six convolutional layers and five pooling layers; The raw edge prediction data is input into the edge prediction neural network model for feature extraction, and the predicted edge image is output. A loss function is constructed based on the material type, and the edge prediction neural network model is optimized based on the loss function to generate a trained edge prediction neural network model.

6. The method of claim 5, wherein: The edge prediction original data is input into the edge prediction neural network model for feature extraction, and a predicted edge image is output. The data output by the seventh convolutional layer is 4 times up-sampled to generate seventh convolutional layer up-sampled data, and the data output by the fourth pooling layer is 2 times up-sampled to generate fourth pooling layer up-sampled data; The seventh convolutional layer up-sampled data, the fourth pooling layer up-sampled data and the data output by the third pooling layer are fused to generate convolutional pooling fusion data; The convolutional pooling fusion data is 8 times up-sampled to output the predicted edge image.

7. A computer device, characterized by A memory and a processor are included, and the memory has stored thereon a computer program capable of being loaded and executed by the processor to perform any one of the methods in claims 1-6.

8. A computer-readable storage medium, characterized in that, A memory has stored thereon a computer program capable of being loaded and executed by the processor to perform any one of the methods in claims 1-6.

Images