Hydrogenation catalyst image recognition method based on improved Center Net

Abstract

The invention belongs to the technical field of image processing, and discloses a hydrogenation catalyst image recognition method based on improved Center Net, and the method comprises the steps: S1, carrying out the sulfuration of a catalyst, transferring the sulfurized catalyst to a TEM, and carrying out the observation of the sulfurized catalyst to complete the image collection of electron microscope equipment; s2, extracting an edge with relatively high contrast by adopting edge detection and a morphological algorithm so as to carry out edge enhancement; s3, Resnet50 is adopted as a backbone network for feature extraction, a space separable convolution kernel is adopted for multi-scale feature extraction of the catalyst edge, an attention module is introduced to adaptively adjust the weight, and target features of the catalyst are highlighted; the FPN is added to the backbone network for deep and shallow feature fusion, and the bonding area of the stripes is segmented to obtain an FPNCNet (improved Center Net); and S4, carrying out image recognition by using the FPNCNet. The method is superior to a traditional catalyst recognition network model, the improved model has good feature extraction ability and generalization ability, and the method has a certain guiding effect on hydrogenation catalyst image recognition work.

Description

Technical Field

[0001] The invention belongs to the technical field of image processing, and in particular relates to a hydrogenation catalyst image recognition method based on improved CenterNet. Background Art

[0002] With the development and widespread application of information technology such as personal computers, the Internet and other communication technologies, human society has entered the information age, and the digitization of information has become the most effective method for information transmission, processing and storage. Digital images have also made great progress in the field of electron microscopy. It is an inevitable development trend of modern electron microscope image processing to convert various images obtained by electron microscope into digital signals in real time for processing and analysis. The transmission electron microscope method can clearly give the active phase information of hydrogenation catalysts, and the statistical results of the average number of layers of active phase crystals have a good correlation with the activity of the catalyst. After accurate identification and segmentation of the electron microscope image of the hydrogenation catalyst, it can be grouped and layered, and the catalyst activity can be judged by the distribution information of the catalyst layer number.

[0003] Traditional methods for determining catalyst activity include conversion frequency method, reaction rate method, activation energy and conversion rate, etc. However, most of the above determination methods have the limitations of difficulty in determination, requiring reaction conversion rate not too high and inaccuracy. Therefore, a method of converting electron microscope images into digital signals for processing and analysis has emerged. Based on this, the present invention proposes a hydrogenation catalyst image recognition method based on improved CenterNet. The improved morphological method is combined with the gradient threshold segmentation method, which can better retain the catalyst information in the image, improve the segmentation accuracy, and provide conditions for determining the key parameters of the catalyst in the future. Summary of the invention

[0004] In order to overcome the above technical problems, the present invention provides a hydrogenation catalyst image recognition method based on improved CenterNet, which solves the problem of low recognition accuracy of hydrogenation catalyst images due to adhesion, stacking and other reasons.

[0005] The present invention adopts the following technical solutions:

[0006] A hydrogenation catalyst image recognition method based on improved CenterNet comprises the following steps:

[0007] S1. After the catalyst is sulfided, it is transferred to a TEM for observation to complete image acquisition of the electron microscope device;

[0008] S2. Use edge detection and morphological algorithms to extract edges with strong contrast for edge enhancement;

[0009] S3. Resnet-50 is used as the backbone network for feature extraction, spatially separable convolution kernels are used for multi-scale feature extraction of catalyst edges, and attention modules are introduced to adaptively adjust weights to highlight the target features of catalysts; FPN (Feature Pyramid Network) is added to the backbone network for deep and shallow feature fusion and segmentation of the bonding area of ​​stripes to obtain FPNCNet (improved CenterNet);

[0010] S4. Image recognition using FPNCNet.

[0011] Preferably, the feature maps of L1, L2 and L3 are firstly subjected to the S-ASPP module to adjust the sampling rate and increase the receptive field, and then the features of different scales are fused together by connecting the 1x1 convolution with the input of the latter layer.

[0012] Preferably, during the upsampling stage, the feature pyramid network uses shallow and deep features to improve the accuracy of the network's detection of small objects.

[0013] Preferably, a CA attention mechanism is introduced in the L2, L3, and L4 layers to capture cross-channel, direction-aware, and position-sensitive information, suppress redundant feature information, and improve feature extraction efficiency, so that the model can more accurately locate and identify objects of interest.

[0014] Preferably, the final feature maps are merged using 3x3 convolution to eliminate the overlapping effect caused by the upsampling process, thereby generating the final feature map to complete the detection task.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] The hydrogenation catalyst image recognition method based on the improved CenterNet of the present invention is superior to the traditional catalyst recognition network model. The improved model has good feature extraction and generalization capabilities, and plays a certain guiding role in the recognition of hydrogenation catalyst images. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a network structure diagram of the present invention;

[0018] Figure 2 It is a catalyst electron microscope image data set designed to verify the method provided by the present invention;

[0019] Figure 3 It is the classification experiment result of the test set image. DETAILED DESCRIPTION

[0020] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The methods in the embodiments are conventional methods in the art unless otherwise specified. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0021] The present invention provides a hydrogenation catalyst image recognition method based on improved CenterNet, comprising the following steps:

[0022] S1: electron microscope equipment image acquisition.

[0023] Before the transmission electron microscope (TEM) analysis, the catalyst should be pre-sulfided. The sulfurization procedure is as follows: take 1g of 80-100 mesh catalyst, place it in a glass reactor, purge it with 50ML / min of N2 for about 10min, then switch to H2 containing 15% H2S, flow rate 100ML / min, heating rate 6℃ / min, rise to 360±30℃, heating rate changed to 1℃ / min, rise to 370±30℃, keep for 4h, then cool to 120±30℃ at a cooling rate of 50℃ / min, switch to N2 purge for 0.5h, then pour the sulfurized catalyst into cyclohexane under N2 protection to prevent oxidation, grind for 40min, and put the catalyst into a sample bottle together with cyclohexane.

[0024] Before dispersing the sample onto the sample holder, the suspension in the sample bottle is first dispersed by ultrasonic vibration, and then an appropriate amount of sample is taken with a dropper and dropped onto the microgrid. After the reasonable experimental parameters are determined, the sample is quickly transferred to the TEM for observation after the solvent evaporates.

[0025] S2: Edge enhancement.

[0026] Since each curve in the image is a black edge line with strong contrast, it is key to extract the edge with strong contrast. The present invention uses edge detection and morphological algorithms to extract edges with strong contrast. The algorithm selects curves that meet the threshold conditions according to constraints such as the shortest pixel, contrast value, length ratio, angle difference, distance, etc., and groups and layers the curves that meet the conditions.

[0027] The basic principle of mathematical morphology is to explore the structural elements on the target image, remove irrelevant shapes such as noise points and burrs, retain the main shape information of the image, and obtain the connection between the various parts of the image and the overall structure.

[0028] Morphological image processing is performed by moving the structural elements in the image and performing a convolution-like method, where the structural elements have arbitrary sizes and contain a combination of 0 and 1. A specific logical operation is performed between each pixel position and the binary image below it, and the result is output at the position corresponding to the pixel in the image. Erosion and dilation are the most basic morphological operations in mathematical morphology, and some other morphological operations, such as opening and closing, are composed of these two basic operations. Assuming f(x,y) is the initial image and g(i,j) is the structural element, erosion and dilation can be defined as respectively.

[0029]

[0030]

[0031] Erosion is used to eliminate small and meaningless objects, and dilation is used to fill holes on objects. The opening and closing operations are based on the sequential processing of erosion and dilation. The opening operation is performed by image erosion and then image dilation, and the closing operation is performed by image dilation and then image erosion. The opening and closing operations are defined as follows.

[0032]

[0033]

[0034] The opening operation can smooth the outline of an image, breaking up narrow discontinuities and removing fine protrusions. The closing operation can remove narrow discontinuities, thin gaps and small holes, and can fill in cracks within the outline.

[0035] Gabor filter can extract the texture features and edge features of the image very well, and can also reduce the impact of illumination changes and noise in the image to a certain extent.

[0036] As an important time-frequency analysis method, Gabor transform is a convolution of a sine wave and a Gaussian kernel function in the spatial domain and a shifted Gaussian function in the frequency domain, where the kernel function is defined as

[0037]

[0038] where x′ and y′ are determined by the following formula

[0039]

[0040] In the above formula, x, y represent the pixel coordinate position, λ represents the wavelength of the cosine function, θ represents the direction factor of the parallel stripes, σ represents the standard deviation of the Gaussian function, and γ represents the spatial aspect ratio. Controls the phase shift of the cosine function.

[0041] The input image contour with a certain width is gradually reduced to a width of only one pixel by removing the edges one by one, thus realizing the skeleton extraction. The skeleton of a circle is its center, and the skeleton of a straight line and an isolated point is itself.

[0042] First, find the 8-point area centered on the boundary point, record the center point as P1, and the 8 points in its area clockwise around the center point as P2, P3, ..., P9, where P2 is above P1. Follow the steps below to complete the skeleton line extraction:

[0043] (1) First, mark the boundary points that meet the following conditions at the same time:

[0044] 1)2≤N(P1)≤6;

[0045] 2) S(P1) = 1;

[0046] 3) P2*P4*P6=0;

[0047] 4) P4*P6*P8=0;

[0048] Where N(P1) is the number of non-zero points in P1, and S(P1) is the number of times the values ​​of these points change from 0 to 1 when the order is P2, P3, ..., P9.

[0049] (2) Same as step (1), except that condition 3) is changed to P2*P4*P8=0; condition 4) is changed to P2*P6*P8=0; after all boundary points have been checked, all marked points are removed.

[0050] (3) Steps (1) and (2) constitute one iteration until no points satisfy the marking condition. At this time, the region formed by the remaining points is the refined skeleton.

[0051] S3: Build FPNCNet, S4. Use FPNCNet for image recognition:

[0052] This method is based on the Resnet-50 backbone network. The improved CenterNet network is named FPNCNet. The network structure is as follows: Figure 1As shown. ResNet-50 solves the gradient vanishing and exploding problems in deep network training by introducing residual connections. Spatially separable convolution is to decompose the standard convolution operation into multiple smaller convolution kernels in the spatial dimension, and use spatially separable convolution kernels to extract the multi-scale features of catalyst stripes. In addition, in order to effectively segment the bonding area of ​​the stripes, FPN (Feature Pyrami dNetwork) is added to the backbone network to achieve deep and shallow feature fusion. The introduction of the attention module to adaptively adjust the weights can effectively highlight the target features of the catalyst. The problem of different linings in images of hydrogenation catalysts of different thicknesses is manifested as large image background noise, a large number of catalysts, small individuals, different lengths, and serious accumulation. First, in order to enhance feature propagation and improve the ability of the network, the feature maps of L1, L2, and L3 are first adjusted by the S-ASPP module to adjust the sampling rate and increase the receptive field, and then the features of different scales are fused together by connecting the 1x1 convolution with the input of the back layer. The network structure introduced here is named DeN. In the upsampling stage, the feature pyramid network uses shallow and deep features to improve the accuracy of the network's detection of small objects. In addition, the CA attention mechanism is introduced in the L2, L3, and L4 layers to capture cross-channel, direction-aware, and position-sensitive information, suppress redundant feature information, and improve feature extraction efficiency, so that the model can more accurately locate and identify objects of interest. Finally, the final feature map is merged using 3x3 convolution to eliminate the overlapping effect caused by the upsampling process, thereby generating the final feature map to complete the detection task.

[0053] The hydrogenation catalyst image recognition method based on the improved CenterNet is verified below.

[0054] exist Figure 2 In the paper, 6000 hydrogenation catalyst images were collected by transmission electron microscopy and used as a dataset. The dataset used for network training was randomly divided into a training set and a test set in a ratio of 5:1. In order to compare the recognition accuracy under different preprocessing methods, the batch size was set to 64 and image enhancement was used. HE means processing using color histogram; Original means original image; FST means image enhancement processing using fuzzy set theory. Since FST will lose edge details, the final classification accuracy will be lower than that of the original image. Since HE can make up for this defect and achieve the highest classification accuracy at the same time, the HE processing results were used in subsequent experiments.

[0055] exist Figure 3In the figure, the classification experiment results of the test set images are shown. According to the operation results of different parameters in Figures (a) to (d), we can see some trends. By increasing the shortest pixel length Lengthmin of the curve to be tested, some short curves can be eliminated. If the contrast value ValueC of the curve to be tested is increased, the curves with weaker contrast can be obviously screened out. Increasing Lengthmin and ValueC will lead to an increase in the false detection rate. Although the false detection rate will decrease, the accuracy will increase and the recall rate will decrease. Therefore, the F1 value is used to evaluate the precision and recall rate as a whole, and when Lengthmin=24 and ValueC=0.7, the F1 value of each indicator is the largest and relatively balanced. Compared with various traditional algorithms, the algorithm proposed in the present invention has greatly improved in indicators such as recognition accuracy and recall rate. The missed detection rate is low, which meets the engineering requirements. FPNCNet may miss too many superimposed fringe phases, so it is necessary to effectively segment the cohesive and superimposed fringe phases to improve the algorithm's ability to identify fringe phases.

[0056] Although the embodiments of the present invention have been shown and described, those skilled in the art will appreciate that various changes, modifications, substitutions and variations may be made to the above embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A hydrogenation catalyst image recognition method based on improved CenterNet, characterized in that: The following steps are involved: S1. After the catalyst is sulfided, it is transferred to a TEM for observation to complete image acquisition of the electron microscope device; S2. Use edge detection and morphological algorithms to extract edges with strong contrast for edge enhancement; S3. Resnet-50 is used as the backbone network for feature extraction, spatially separable convolution kernels are used for multi-scale feature extraction of catalyst edges, and attention modules are introduced to adaptively adjust weights to highlight the target features of the catalyst; Add FPN (Feature Pyramid Network) to the backbone network for deep and shallow feature fusion and segmentation of the bonding area of ​​the stripes to obtain FPNCNet (improved CenterNet); S4. Image recognition using FPNCNet.

2. The hydrogenation catalyst image recognition method based on improved CenterNet according to claim 1, characterized in that: The feature maps of L1, L2, and L3 are firstly processed through the S-ASPP module to adjust the sampling rate and increase the receptive field, and then the features of different scales are fused together by connecting the 1x1 convolution with the input of the latter layer.

3. The hydrogenation catalyst image recognition method based on improved CenterNet according to claim 2, characterized in that: In the upsampling stage, the Feature Pyramid Network uses shallow and deep features to improve the network's accuracy in detecting small objects.

4. The hydrogenation catalyst image recognition method based on improved CenterNet according to claim 3, characterized in that: The CA attention mechanism is introduced in the L2, L3, and L4 layers to capture cross-channel, direction-aware, and position-sensitive information, suppress redundant feature information, and improve feature extraction efficiency, so that the model can more accurately locate and identify objects of interest.

5. The hydrogenation catalyst image recognition method based on improved CenterNet according to claim 4, characterized in that: The final feature maps are merged using 3x3 convolution to eliminate the overlapping effect caused by the upsampling process, thereby generating the final feature map to complete the detection task.

Images