CN112967271B - A Casting Surface Defect Recognition Method Based on Improved DeepLabv3+ Network Model - Google Patents

Abstract

The invention provides a casting surface defect identification method based on an improved deep Labv3+ network model, which comprises the following steps: s1, collecting a casting image data set to obtain a training set and a test set; step S2, constructing a network model, and performing data training and correction on the network model through a training set and a testing set to generate a defect detection network; step S3, designing a loss function of the defect detection network; and step S4, the defect detection network identifies and outputs the defect detection result of the casting and displays the detection duration. According to the invention, the surface defects of the castings are identified by adopting a deep learning method, so that the accuracy and speed of defect identification are improved, and a new idea is provided for quality detection of industrial castings.

Description

A Casting Surface Defect Recognition Method Based on Improved DeepLabv3+ Network Model

technical field

The invention relates to the technical field of industrial casting defect detection, in particular to a casting surface defect identification method based on an improved DeepLabv3+ network model.

Background technique

In industrial casting production, various types of castings are widely used. However, in the process of production, castings still inevitably have defects such as cracks and shrinkage holes. These defects can seriously affect the quality of finished castings, which in turn affects the quality of the entire manufactured product. Efficient defect detection technology is one of the important links to ensure the quality of casting products. Therefore, in the period of rapid development and production of the modern foundry industry, accurate and efficient casting surface defect detection technology is an important development direction of industrial product detection.

At present, most of the domestic foundry production still adopts the method of manual sampling inspection to detect casting defects, mainly relying on the inspection personnel to observe with the naked eye and subjectively judge to complete the inspection. This method relies on the prior knowledge of the detection personnel, has strong subjectivity, lacks accuracy and standardization, and cannot guarantee the efficiency. Other industrial casting surface defect detection methods include eddy current coil inspection and ultrasonic inspection. However, these traditional methods all have different drawbacks. The detection work consumes a lot of manpower and material resources, and the final detection results also require manual processing to make judgments. At the same time, during ultrasonic inspection and eddy current coil inspection, there may be contact with the surface defects of the casting, and physical and chemical changes may occur, further expanding the defect area, which is not conducive to the improvement of inspection accuracy.

SUMMARY OF THE INVENTION

In view of the above problems, in order to improve the current casting defect detection dilemma, the present invention proposes a casting surface defect identification method based on the improved DeepLabv3+ network model. Automatic real-time online and intelligent detection of casting surface defects.

The collected workpiece surface defect images are made into a defect dataset and sent to the improved neural network, and then trained and learned through a self-defined network based on semantic segmentation, and finally the trained network is obtained for image detection and marking of workpiece surface defects. The defect area is then combined with the defect detection software to output the detection time. At the same time, the network model can be replaced to compare different models. Under the research of this method, the defect detection of the workpiece can be intelligently identified, so as to achieve high-precision detection and reduce manual intervention. .

In order to achieve the above purpose, a method for identifying surface defects of castings based on the improved DeepLabv3+ network model provided by the present invention includes the following steps:

Step S1, collecting a casting image data set to obtain a training set and a test set: the industrial camera collects the casting image, Labelme marks the casting defects, generates an image data set, and divides the image data set into a training set and a test set;

Step S2, constructing a network model, and performing data training and correction on the network model through the training set and the test set to generate a defect detection network;

Step S3, designing the loss function of the defect detection network: the loss function is a function including the cross-entropy loss of the prediction result and the cross-entropy loss of the real value;

Step S4, the defect detection network identifies and outputs the casting defect detection result, and displays the detection duration.

Preferably, step S2 specifically includes the following steps:

Step S21, constructing and improving the DeepLabV3+ network model: improving the encoding and decoding module in the existing DeepLabV3+ network model, the encoding and decoding module adopts a depth separable convolutional network, and the number of channels is deleted;

Step S22, perform data training and correction to generate a defect detection network: use the improved DeepLabv3+ network model as a network model for casting surface defect identification to extract the casting surface defect features of the originally collected casting images, and improve the DeepLabv3+ network by using the training data in the training set. The model is trained, and then the test set is input into the improved DeepLabv3+ network model for correction until the prediction accuracy of the generated defect detection network meets the prediction accuracy threshold.

Preferably, in the step S4, outputting the detection result of the casting is specifically realized by outputting the detection result through a visualization software facing the defect detection of the industrial casting.

Preferably, in the step S21, the encoding and decoding module includes an encoding module and a decoding module, the encoding module is connected to the decoding module, and the encoding module includes a deep convolutional network layer, an encoder and a hole space pyramid pool The deep convolutional network layer is the backbone model, and the low-level semantic features including defect shape information obtained after convolution of the input picture through the deep convolutional network layer are directly transmitted to the decoding module, and the input picture is passed through The high-level semantic features including defect location categories obtained by the encoder and the deep convolutional network layer are passed into the hole space pyramid pooling module, and the hole space pyramid pooling module includes a hole convolution module and a four-layer convolution layer. and one layer of pooling layer, the four layers of convolution layers are connected in series with the one layer of pooling layers, and the hole convolution modules respectively use hole convolution blocks with different expansion rates to control the resolution of the features extracted by the encoder. rate, respectively perform hole convolution to obtain four feature maps of the four-layer convolution layer, and the hole convolution module obtains a feature map of the one-layer pooling layer after pooling.

Preferably, for the low-level semantic feature information obtained by the backbone model, the encoding module operates on the low-level semantic feature through a convolution to reduce the number of channels, and obtains the final semantic information of the low-level semantic feature after the operation, and transmits it to the In the decoder of the decoding module; for the high-level semantic features obtained by the hole space pyramid pooling module, the decoding module obtains the semantic information of the final high-level semantic features through one upsampling, and is passed to the decoder as the output result. middle.

Preferably, the encoding module further includes a stack layer, a channel adjustment convolution layer, an intermediate feature layer, an intermediate feature layer stack layer, a post-processing convolution layer and a post-processing module, and the stack layer is the four-layer convolution layer and the feature map obtained by the layer of pooling layer are stacked to form, the stacked layer is convolved to adjust the channel, the channel adjustment convolution layer is formed, and the channel adjustment convolution layer is adjusted in length and width , the intermediate feature layer is formed, the intermediate feature layer stack is formed after the intermediate feature layer is stacked, the intermediate feature layer stack is convoluted to form the post-processing convolution layer, and then the post-processing convolution layer is formed. The convolutional layer is post-processed by the post-processing module.

Preferably, the post-processing module includes a pixel point classification module and a pixel point category probability calculation module, the pixel point classification module classifies each pixel point on the image of the post-processing convolution layer, and the pixel point category The probability calculation module uses softmax to solve the probability of each pixel category.

Preferably, the decoding module upsamples the high-level semantic features output by the encoder using bilinear upsampling with a sampling factor of 4, and the input of the decoding module consists of two parts, one part comes from the backbone model of the The direct output of the deep convolutional network layer is connected to the corresponding low-level semantic features with the same spatial resolution output from the backbone model, and the other part is imported into the hole spatial pyramid pooling module through the deep convolutional network layer of the backbone model. , find a low-level semantic feature map with the same resolution as the semantic feature output by the encoder, reduce the number of channels through 1*1 convolution to make it equal to the channel proportion of the semantic feature output by the encoder, and decode the The low-level semantic features and low-level semantic feature maps obtained from the input of the two parts of the module are connected together. After the connection is processed, it is refined by a 3*3 thinning convolution, and then the bilinear sampling factor is 4. Sampling to get the final prediction result.

Preferably, the upsampling adopted by the output of the encoding module to the decoding module and the upsampling adopted by the output of the decoding module are bilinear interpolation, and the definition formula is as follows:

src _x =des _x *src _ω /des _ω

src _y =des _y *src _h /des _h

In the formula, src _x represents the x-coordinate of the pixel in the original image, src _y represents the y-coordinate of the pixel in the original image, des _x represents the x-coordinate of the pixel in the target image, des _y represents the y-coordinate of the pixel in the target image, and src _ω represents The width of the original image, src _h is the height of the original image, des _ω is the width of the target image, and des _h is the height of the target image.

Preferably, the loss function in the step S3 uses a cross-entropy loss function, and the formula is as follows:

J=-[y·log(p)+(1-y)·log(1-p)]

Where y represents the label of the sample, the positive class is 1, and the negative class is 0; p represents the probability that the sample is predicted to be positive.

The present invention can achieve the following beneficial effects: the improved DeepLabv3+ network model based on the present invention can be used for industrial casting defect identification, and the identification network can be directly applied to the field of industrial casting defect detection, and digital defect data for defect identification is generated after identification, The result storage and query can correspond one-to-one with the casting information, providing a certain data feedback, which can be used for the optimization of the casting process and provide guidance for the subsequent process. In particular, it is pointed out that this method mainly targets the grinding defects of industrial castings , for casting defect detection and identification.

The network model based on the improved DeepLabV3+ adopted in the present invention has the advantages of fast recognition speed and high recognition accuracy, and transforms the traditional defect detection process into an excellent defect detection and identification method based on deep learning algorithm and deep neural convolution network, which is effective Solve the shortcomings of traditional craftsmanship.

The invention designs a visualization software for casting defect detection. The software can select the defect pictures to be detected by setting the network model of defect detection, and output the detection results and detection time after the detection is completed. The visualization results are clear, and the software operation is simple and easy to use. , which greatly improves the current situation that the current casting defect detection has the drawbacks of manual participation.

Description of drawings

Fig. 1 is the flow chart of a kind of casting surface defect identification method based on improved DeepLabv3+ network model of the present invention;

2 is a schematic diagram of the overall network structure of a preferred embodiment of a method for identifying surface defects of castings based on an improved DeepLabv3+ network model of the present invention;

Fig. 3 is the schematic diagram of the coding module of the overall network structure shown in Fig. 2;

4 is a schematic diagram of a decoding module of the overall network structure shown in FIG. 3;

5 is a schematic diagram of a backbone model network structure of a preferred embodiment of a method for identifying surface defects of castings based on an improved DeepLabv3+ network model of the present invention;

6 is a schematic diagram of a recognition result of a preferred embodiment of a method for recognizing casting surface defects based on an improved DeepLabv3+ network model of the present invention;

7 is a schematic diagram of a Qt design software interface of a preferred embodiment of a method for identifying surface defects of castings based on an improved DeepLabv3+ network model of the present invention.

Detailed ways

In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, the following will be described in detail with reference to the accompanying drawings and specific embodiments.

Aiming at the existing problems, the present invention provides a casting surface defect identification method based on an improved DeepLabv3+ (a semantic segmentation model) network model. The invention is a detection method based on a neural network, and by improving the DeepLabv3+ network model to detect casting defects, it can realize automatic real-time online and intelligent detection of casting surface defects. The collected workpiece surface defect images are made into a defect dataset and sent to the improved neural network, and then trained and learned through a self-defined network based on semantic segmentation, and finally the trained network is obtained for image detection and marking of workpiece surface defects. The defect area is then combined with the defect detection software to output the detection time. At the same time, the network model can be replaced to compare different models. Under the research of this method, the defect detection of the workpiece can be intelligently identified, so as to achieve high-precision detection and reduce manual intervention. .

As shown in Figure 1, a method for identifying surface defects of castings based on the improved DeepLabv3+ network model of the present invention includes the following steps:

Step S1, collecting a casting image data set to obtain a training set and a test set: the industrial camera collects casting images, Labelme (a deep learning labeling tool) marks casting defects, generates an image data set, and divides the image data set into training set and test set set;

Define the standard of casting defects, and use Labelme to mark the defect area of each image. The marked images include defect location and defect area, forming a data set of common industrial casting surface defects. The data set includes a total of 244 images. The image pixel size is 1408*1580.

Each defect image in the casting defect dataset contains only one type of defect, and the size of the image is uniformly adjusted to 1408*1580 pixels. Further, the obtained casting data set is divided into a training set and a test set. The training set is the pictures that have been marked with defects, a total of 211 pictures, and the test set is the unlabeled pictures of casting defects, a total of 33 pictures.

Step S2, constructing a network model, and performing data training and correction on the network model through the training set and the test set to generate a defect detection network;

Step S3, designing the loss function of the defect detection network: the final loss function is a function including the cross-entropy loss of the prediction result and the cross-entropy loss of the real value;

Step S4, the defect detection network identifies and outputs the casting defect detection result, and displays the detection duration.

Step S2 specifically includes the following steps:

Step S21, constructing and improving the DeepLabV3+ network model: improving the encoding and decoding module in the existing DeepLabV3+ network model, the encoding and decoding module adopts a depth separable convolutional network, and the number of channels is deleted;

The improved DeepLabv3+ network model is used to build the convolutional neural network structure for training. The backbone model and the encoder-decoder module in the DeepLabV3+ network model are improved. However, the number of parameters is too large and the calculation time is too long. Therefore, this method improves the encoding and decoding module, and replaces the ordinary convolution of the encoding and decoding part and the decoding part with a depthwise separable convolution. The number of channels has also undergone a certain amount of pruning.

The backbone network uses the Xception (a kind of network convolution structure) network structure. In order to speed up the network calculation speed and reduce the detection time of casting defect pictures, this method replaces the Xception network with the MobileNet V2 (a lightweight convolutional neural network) network. .

Step S22, perform data training and correction to generate a defect detection network: use the improved DeepLabv3+ network model as a network model for casting surface defect identification to extract the casting surface defect features of the originally collected casting images, and improve the DeepLabv3+ network by using the training data in the training set. The model is trained, and then the test set is input into the improved DeepLabv3+ network model for correction until the prediction accuracy of the generated defect detection network meets the prediction accuracy threshold. The purpose of obtaining a good network prediction model is achieved by adjusting the training parameters during training.

The relevant parameters of the training network settings are as follows: the learning rate is set to 0.01, the number of threads for loading data (batch) is set to 4, the number of training steps is set to 40, and the batch-size is set to 4, using Intel I7 CPU, GPU: Nvidia_1080TI.

As shown in FIG. 2 , FIG. 3 and FIG. 4 , in step S21 , the encoding and decoding module includes an encoding module and a decoding module, the encoding module is connected to the decoding module, and the encoding module includes a deep convolutional network Layer, encoder and hole space pyramid pooling module, the deep convolutional network layer is the backbone model, the low-level semantic features obtained after convolution are passed into the decoding module, and the high-level semantic features obtained by the deep convolutional network layer are Semantic features are passed into the hole spatial pyramid pooling module.

In the step S21, the encoding and decoding module includes an encoding module and a decoding module, the encoding module is connected to the decoding module, the encoding module includes a backbone model and a hole space pyramid pooling module, and the backbone model is a depth. In the convolutional network layer MobileNet V2, the low-level semantic features including the shape information of defects obtained after the input picture is convolved through the backbone model are directly passed into the decoding module, and the high-level semantics including the defect location category are obtained from the input picture through the encoder Features are passed into the hole space pyramid pooling module. FIG. 5 is a schematic diagram of the network structure of the backbone model.

The picture with defects is passed into the encoding module, and through the deep convolutional network layer, the deep neural network model contains the convolution of the serial structure, the obtained low-level semantic features are passed into the decoding module, and the high-level semantic features are passed to the decoding module. into the hole space pyramid pooling module.

The low-level semantic features including defect shape information obtained after the input picture is convolved through the deep convolutional network layer are directly passed into the decoding module, and the input picture obtained through the encoder and the deep convolutional network layer includes defects. The high-level semantic features of the location category are passed into the hole space pyramid pooling module, and the hole space pyramid pooling module includes a hole convolution module, a four-layer convolution layer, and a pooling layer. The four-layer convolution layer It is serially connected with the one-layer pooling layer, and the atrous convolution module adopts atrous convolution blocks with different expansion rates respectively, controls the resolution of the feature extracted by the encoder, and performs atrous convolution respectively to obtain the four-layer volume. The four feature maps of the accumulation layer are obtained after the atrous convolution module performs pooling to obtain a feature map of the one-layer pooling layer.

For the low-level semantic feature information obtained by the backbone model, the encoding module operates on the low-level semantic feature through a convolution to reduce the number of channels, and after the operation obtains the final semantic information of the low-level semantic feature, which is sent to the decoding module In the decoder of ; for the high-level semantic features obtained by the hole space pyramid pooling module, the decoding module obtains the semantic information of the final high-level semantic features through one upsampling, and transmits it to the decoder as an output result.

The output of the bottom layer is processed by hole convolution, and the number of channels is adjusted by 1*1 convolution to obtain the b0 layer. The hole convolution block with an expansion rate of 18 gets the b3 layer. Finally, the four convolution layers and one pool are stacked with 1*1 convolution for channel adjustment, and then the length and width of the intermediate feature layer are processed. The intermediate feature layers are stacked, the final processing is performed with 3*3 convolution, and then each pixel is classified, and finally the probability of each pixel category is calculated by softmax.

Based on the improved DeepLabv3+ network model, an encoding-decoding structure model using atrous convolution is constructed. In this encoding-decoding architecture, the resolution of the features extracted by the controllable encoder is introduced, and the atrous convolution is used to balance the accuracy and time-consuming. . The atrous convolution used can increase the receptive field by cross-pixel extraction during feature extraction without losing information, so that each convolution output contains a large range of information. Atrous convolution is used in Feature extraction is performed in the encoding module.

When the output of the encoding module with stride of 16 is output, a balance of rate and accuracy is achieved. When the output step size is 8, the coding module outputs, the precision is higher, but the computational complexity increases. After balancing the advantages of various aspects, this method decides to use an encoding module with an output step size of 16.

The high-level semantic features are encoded and decoded by the hole pyramid pooling module, and are convolved and pooled with the four-layer convolutional layers of the four hollow convolutional layers and the one-layer pooling layer of one pooling layer respectively. , to obtain five feature maps, which are further connected into a five-layer empty space pyramid pooling module.

However, since the corresponding low-level semantic features contain more channel information, it may exceed the output encoding features and cause training difficulties. Therefore, before the connection operation, the encoding module operates on the low-level semantic features through a 1*1 convolution. In order to reduce the number of channels, the final semantic information is obtained after the operation, and the output result obtained by the decoding module through one upsampling is passed to the decoder.

The encoding module further includes a stacking layer, a channel adjustment convolutional layer, an intermediate feature layer, an intermediate feature layer stacking layer, a post-processing convolutional layer, and a post-processing module, where the stacking layer is the four-layer convolutional layer and the The feature maps obtained by a layer of pooling layers are stacked to form, the stacked layers are convoluted to adjust the channel, the channel adjustment convolution layer is formed, and the channel adjustment convolution layer is adjusted in length and width to form the channel adjustment convolution layer. The intermediate feature layer is stacked to form the intermediate feature layer stack, the intermediate feature layer stack is convoluted to form the post-processing convolution layer, and then the post-processing convolution layer is formed. Post-processing is performed by the post-processing module.

The post-processing module includes a pixel point classification module and a pixel point category probability calculation module, the pixel point classification module classifies each pixel point on the image of the post-processing convolution layer, and the pixel point category probability calculation module. Use softmax (an excitation function of the output layer) to solve the probability of each pixel category.

The four-layer convolutional layer and the one-layer pooling layer are serially connected and then subjected to a 1*1 convolution operation and then transmitted to the decoding module as high-level semantic feature information. The deep neural network layers of the backbone model in the encoding module are passed into the decoding module together.

The result of the 1*1 convolution operation on the high-level semantic features passed into the decoding module by the encoding module and the low-level semantic feature information from the backbone model are connected together to form a parallel structure, and then a 3*3 convolution is performed. Operation, through a quadruple upsampling, the decoding module is further formed, and finally the prediction result is sent out.

Specifically, the decoding module upsamples the high-level semantic features output by the encoder using bilinear upsampling with a sampling factor of 4, and the input of the decoding module consists of two parts, one part comes from the backbone model of the The direct output of the deep convolutional network layer is connected to the corresponding low-level semantic features with the same spatial resolution output from the backbone model, and the other part is imported into the hole spatial pyramid pooling module through the deep convolutional network layer of the backbone model. , find a low-level semantic feature map with the same resolution as the semantic feature output by the encoder, reduce the number of channels through 1*1 convolution to make it equal to the channel proportion of the semantic feature output by the encoder, and decode the The low-level semantic features and low-level semantic feature maps obtained from the input of the two parts of the module are connected together. After the connection is processed, it is refined by a 3*3 thinning convolution, and then the bilinear sampling factor is 4. Sampling to get the final prediction result.

The upsampling adopted by the output of the encoding module input to the decoding module and the upsampling adopted by the output of the decoding module are both bilinear interpolation, and the definition formula is as follows:

src _x =des _x *src _ω /des _ω

src _y =des _y *src _h /des _h

In the formula, src _x represents the x-coordinate of the pixel in the original image, src _y represents the y-coordinate of the pixel in the original image, des _x represents the x-coordinate of the pixel in the target image, des _y represents the y-coordinate of the pixel in the target image, and src _ω represents The width of the original image, src _h is the height of the original image, des _ω is the width of the target image, and des _h is the height of the target image.

In the step S3, the loss function uses the cross-entropy loss function, and the formula is as follows:

J=-[y·log(p)+(1-y)·log(1-p)]

Where y represents the label of the sample, the positive class is 1, and the negative class is 0; p represents the probability that the sample is predicted to be positive.

The outputting of the casting detection results in the step S4 is specifically to realize the outputting of the detection results through visualization software for industrial casting defect detection.

As shown in Figure 6 and Figure 7, this software inputs images of industrial castings with defects. Different recognition models can be added, and the final image of defect detection results is output. The operation is simple and the visualization results are clear. Users can directly detect and identify defects. Casting pictures. Visualization software is a software for industrial casting defect detection. It is a visual industrial software based on Qt. It has the characteristics of simple operation, clear visualization, and displaying the detection time of defect detection and optional detection network.

The present invention can achieve the following beneficial effects: the improved DeepLabv3+ network model based on the present invention can be used for industrial casting defect identification, and the identification network can be directly applied to the field of industrial casting defect detection, and digital defect data for defect identification is generated after identification, The result storage and query can correspond one-to-one with the casting information, providing a certain data feedback, which can be used for the optimization of the casting process and provide guidance for the subsequent process. In particular, it is pointed out that this method mainly targets the grinding defects of industrial castings , for casting defect detection and identification.

The network model based on the improved DeepLabV3+ adopted in the present invention has the advantages of fast recognition speed and high recognition accuracy, and transforms the traditional defect detection process into an excellent defect detection and identification method based on deep learning algorithm and deep neural convolution network, which is effective Solve the shortcomings of traditional craftsmanship.

The invention designs a visualization software for casting defect detection. The software can select the defect pictures to be detected by setting the network model of defect detection, and output the detection results and detection time after the detection is completed. The visualization results are clear, and the software operation is simple and easy to use. , which greatly improves the current situation that the current casting defect detection has the drawbacks of manual participation.

The above are the preferred embodiments of the present invention. It should be pointed out that for those skilled in the art, without departing from the principles of the present invention, several improvements and modifications can be made. It should be regarded as the protection scope of the present invention.

Claims (5)

1. a casting surface defect identification method based on improving DeepLabv3+ network model, is characterized in that, comprises the steps: Step S1, collecting a casting image data set to obtain a training set and a test set: the industrial camera collects the casting image, Labelme marks the casting defects, generates an image data set, and divides the image data set into a training set and a test set; Step S2, constructing a network model, and performing data training and correction on the network model through the training set and the test set to generate a defect detection network; Step S3, designing the loss function of the defect detection network: the loss function is a function including the cross-entropy loss of the prediction result and the cross-entropy loss of the real value; Step S4, the defect detection network identifies and outputs the casting defect detection result, and displays the detection duration; Step S2 specifically includes the following steps: Step S21, constructing and improving the DeepLabV3+ network model: improving the encoding and decoding module in the existing DeepLabV3+ network model, the encoding and decoding module adopts a depth separable convolutional network, and the number of channels is deleted; Step S22, perform data training and correction to generate a defect detection network: use the improved DeepLabv3+ network model as a network model for casting surface defect identification to extract the casting surface defect features of the originally collected casting images, and improve the DeepLabv3+ network by using the training data in the training set. The model is trained, and then the test set is input into the improved DeepLabv3+ network model to correct it, until the prediction accuracy of the generated defect detection network meets the prediction accuracy threshold; In the step S21, the encoding and decoding module includes an encoding module and a decoding module, the encoding module is connected to the decoding module, and the encoding module includes a deep convolutional network layer, an encoder and a hole space pyramid pooling module, The deep convolutional network layer is the backbone model, and the low-level semantic features including defect shape information obtained by convolution of the input picture through the deep convolutional network layer are directly transferred to the decoding module, and the input picture is passed through the encoder and the decoder. The high-level semantic features including defect location categories obtained by the deep convolutional network layer are passed into the hole space pyramid pooling module, and the hole space pyramid pooling module includes a hole convolution module, a four-layer convolution layer and a layer of Pooling layer, the four-layer convolution layer is serially connected with the one-layer pooling layer, and the hole convolution modules respectively use hole convolution blocks with different expansion rates to control the resolution of the features extracted by the encoder, respectively. After the hole convolution is performed, four feature maps of the four-layer convolution layer are obtained, and the hole convolution module is pooled to obtain a feature map of the one-layer pooling layer; For the low-level semantic feature information obtained by the backbone model, the encoding module operates on the low-level semantic feature through a convolution to reduce the number of channels, and after the operation obtains the final semantic information of the low-level semantic feature, which is sent to the decoding module In the decoder; for the high-level semantic features obtained by the hole space pyramid pooling module, the decoding module obtains the semantic information of the final high-level semantic features through one-time upsampling, which is passed into the decoder as an output result; The decoding module upsamples the high-level semantic features output by the encoder using bilinear upsampling with a sampling factor of 4, and the input of the decoding module consists of two parts, one part from the depth convolution of the backbone model. The direct output of the network layer is connected to the corresponding low-level semantic features with the same spatial resolution output from the backbone model, and the other part is imported into the hole spatial pyramid pooling module through the deep convolutional network layer of the backbone model. A low-level semantic feature map with the same resolution as the semantic feature output by the encoder, after 1*1 convolution, the number of channels is reduced to make it equal to the channel proportion of the semantic feature output by the encoder, and the decoding module is divided into two parts. The low-level semantic features and low-level semantic feature maps obtained from the input are connected together. After the connection processing, they are then refined by a 3*3 thinning convolution, and then the final result is obtained by bilinear upsampling with a sampling factor of 4. the forecast result; The upsampling adopted by the output of the encoding module input to the decoding module and the upsampling adopted by the output of the decoding module are both bilinear interpolation, and the definition formula is as follows: src _x =des _x *src _ω /des _ω src _y =des _y *src _h /des _h In the formula, src _x represents the x-coordinate of the pixel in the original image, src _y represents the y-coordinate of the pixel in the original image, des _x represents the x-coordinate of the pixel in the target image, des _y represents the y-coordinate of the pixel in the target image, and src _ω represents The width of the original image, src _h is the height of the original image, des _ω is the width of the target image, and des _h is the height of the target image. 2. a kind of casting surface defect identification method based on improved DeepLabv3+ network model according to claim 1, is characterized in that, described in described step S4, output casting detection result is specifically by the visualization software that faces industrial casting defect detection Realize the output detection result. 3. a kind of casting surface defect identification method based on improving DeepLabv3+ network model according to claim 1, is characterized in that, described coding module also comprises stacking layer, channel adjustment convolution layer, middle feature layer, middle feature layer stacking layer, post-processing convolutional layer and post-processing module, the stacking layer is formed by stacking the feature maps obtained by the four-layer convolutional layer and the one-layer pooling layer, and convolving the stacked layer to obtain Adjust the channel to form the channel adjustment convolution layer, and adjust the length and width of the channel adjustment convolution layer to form the intermediate feature layer. After the intermediate feature layers are stacked, the intermediate feature layer stack is formed. The post-processing convolution layer is formed after the stacking layers of the intermediate feature layers are convoluted, and then the post-processing convolution layer is post-processed by the post-processing module. 4. a kind of casting surface defect identification method based on improved DeepLabv3+ network model according to claim 3, is characterized in that, described post-processing module comprises pixel classification module and pixel classification probability calculation module, and described pixel classification The module classifies each pixel point on the image of the post-processing convolution layer, and the pixel point category probability calculation module uses softmax to calculate the probability of each pixel point category. 5. a kind of casting surface defect identification method based on improving DeepLabv3+ network model according to claim 1, is characterized in that, in described step S3, loss function uses cross entropy loss function, and formula is as follows: J=-[y·log(p)+(1-y)·log(1-p)] Where y represents the label of the sample, the positive class is 1, and the negative class is 0; p represents the probability that the sample is predicted to be positive.

Images