Wheat impurity-crushing detection method and system based on improved DeeplabV < 3 + >

Abstract

The invention discloses a wheat impurity-crushing detection method and system based on improved DeeplabV < 3 + >, and relates to the technical field of agricultural equipment intellectualization. The detection system comprises a wheat image acquisition device, an embedded processor and a display device; and the embedded processor is used for processing the static wheat image transmitted by the industrial camera by using an improved DeeplabV3 + algorithm to obtain the impurity rate and the breakage rate of the wheat. According to the method, the DeeplabV3 + algorithm is taken as a baseline, and the expansion prediction, the convolution attention mechanism module CBAM and the detail detection head are introduced, so that the learning ability of the model for the key features of the large-size image is improved, the segmentation result is finer, and the precise recognition of the broken grains which are difficult to recognize is realized.

Description

Technical Field

[0001] The present application belongs to the field of intelligent agricultural equipment technology, and specifically relates to a wheat impurity and breakage detection method and system based on improved DeeplabV3+. Background Art

[0002] Mechanized harvesting has become the main way of harvesting crops. However, a 1% breakage rate during mechanized harvesting by a combine harvester will result in a grain loss of up to 1.377 million tons.

[0003] The breakage rate and impurity rate are important indicators for measuring the quality of wheat combine harvesters. The breakage rate refers to the mass percentage of broken grains in the wheat harvest, and the impurity rate refers to the mass percentage of impurities such as wheat husks and straw in the harvest. According to the industry standard "Technical Conditions for Full-Feed Combine Harvesters JB / T5117-2017", the wheat breakage rate should be ≤1% and the impurity rate should be ≤2%.

[0004] At present, the detection of impurity and breakage rate of wheat in China is still carried out manually, with manual sampling from multiple points of the grain outlet, screening out broken grains and impurities, weighing them, and then calculating the breakage rate and impurity rate. However, manual detection is time-consuming, labor-intensive, costly, and has poor repeatability. In addition, when the sample size is small, the objectivity of the test results is poor, which makes it difficult to meet the testing needs of large quantities of wheat. In recent years, some domestic scholars have developed a combine harvester operation quality detection system, which intelligently controls the working parameters such as the concave plate gap according to information such as the distribution density, moisture content, and crop variety of different wheat fields during the operation of the combine harvester. However, there are still problems such as the detection accuracy and detection speed cannot meet the requirements of field operations. Summary of the invention

[0005] In response to the shortcomings in the prior art, the present application provides a wheat impurity and breakage detection method and system based on the improved DeeplabV3+, which can collect static wheat images during the mechanized wheat harvesting process and perform online analysis of the impurity content and breakage rate, thereby reducing losses and waste during the wheat harvesting process.

[0006] To achieve the above objectives, this application provides the following technical solutions:

[0007] A wheat impurity-breakage detection method based on improved DeeplabV3+, specifically:

[0008] The embedded processor uses the improved DeeplabV3+ algorithm to process the static wheat image transmitted by the industrial camera to obtain the impurity rate and breakage rate of the wheat, which are displayed on the display device.

[0009] Among them, the improved DeeplabV3+ algorithm processes the static wheat image transmitted by the industrial camera, including: the static wheat image is pre-processed by expansion prediction to output a 512 pixel × 512 pixel wheat pre-processed picture, and the wheat pre-processed picture is input into the backbone network Xception-65 to output the deep feature map and the shallow feature map; the convolutional attention mechanism module CBAM and the atrous spatial convolution pooling pyramid ASPP are used to process the deep feature map, the deep feature map and the shallow feature map are decoded and processed, and the segmentation loss value is output; the shallow feature map is processed by the detail detection head to output the detail loss value; the total loss value is calculated using the total loss function, and finally the impurity rate and breakage rate of wheat are calculated using the post-processing of expansion prediction.

[0010] Furthermore, the improved DeeplabV3+ algorithm is used to achieve:

[0011] Use dilation prediction for images processed by DeeplabV3+;

[0012] Add a convolutional attention mechanism module CBAM to the decoding part of DeeplabV3+;

[0013] Add detail detection head part to DeeplabV3+ framework.

[0014] Furthermore, expansion prediction includes two parts: pre-processing and post-processing;

[0015] Among them, the pre-processing of expansion prediction performs overlapping sliding window processing on the static wheat image, outputs 15 512 pixel × 512 pixel wheat pre-processing pictures and splices them into multi-dimensional data and inputs them into the embedded processor, which performs model inference operation on the multi-dimensional data to obtain each segmentation result map;

[0016] The post-processing of expansion prediction splices the central areas of each segmentation result map, and uses the pixel quality mathematical model to calculate the wheat impurity rate and breakage rate based on the pixel distribution of complete grains, broken grains, husks, and straws.

[0017] Furthermore, the impurity rate and breakage rate of wheat are calculated using a pixel quality mathematical model, including:

[0018] The mass of each part is obtained through the correspondence between the pixels and mass of intact kernels, broken kernels, wheat husks, and wheat straws, and the impurity rate and breakage rate of wheat are calculated based on the mass of each part.

[0019] Furthermore, the calculation formula of impurity rate is:

[0020]

[0021] Among them, R impurity is the impurity content, mg is the mass of the whole kernel, m b is the mass of broken kernels, m s is the mass of the straw, m h For the quality of wheat husk.

[0022] Furthermore, the calculation formula of the crushing rate is:

[0023]

[0024] Among them, R broken is the crushing rate, m g is the mass of the whole kernel, m b The mass of broken kernels.

[0025] Furthermore, the convolutional attention mechanism module CBAM includes a channel attention module and a spatial attention module;

[0026] Among them, the channel attention module is used to process the deep feature map, obtain the channel attention weight and multiply the channel attention weight with each channel of the deep feature map;

[0027] The spatial attention module processes the deep feature map output by the channel attention module to obtain the spatial attention weight and multiplies the spatial attention weight with the input feature map of the convolutional attention mechanism module CBAM.

[0028] Furthermore, the detail detection header part is used to implement:

[0029] The shallow feature map passes through a 3×3 convolutional layer, a BN layer, a ReLU activation layer, and a 1×1 convolutional layer in turn to output a detail feature map. At the same time, the Laplacian operator with a step size of 1, 2, and 4 is used to obtain the multi-scale detail information of the label map. The detail feature map is upsampled to the original size, and the predicted details are converted into a binary label edge map through a 1×1 convolution. Finally, the detail feature map and the binary label edge map are combined to calculate the detail loss value.

[0030] Furthermore, the calculation formula of the total loss function is:

[0031] Loss = F s +D s +F d +D d ;

[0032] Among them, Loss is the total loss function, F s The segmentation loss value calculated by Focal Loss, D s The segmentation loss value calculated by Dice Loss, F d The detail loss value calculated for Focal Loss, Dd Detail loss value calculated for Dice Loss.

[0033] A wheat impurity-breakage detection system based on improved DeeplabV3+, comprising a wheat image acquisition device, an embedded processor and a display device, wherein the embedded processor is connected to the wheat image acquisition device and the display device respectively, and the embedded processor is used to process a static wheat image transmitted by an industrial camera using an improved DeeplabV3+ algorithm to obtain the impurity rate and breakage rate of the wheat;

[0034] The wheat impurity-breakage detection system based on improved DeeplabV3+ is used to implement the wheat impurity-breakage detection method based on improved DeeplabV3+ as described in any one of the above

[0035] The beneficial effects of this application are:

[0036] (1) Based on the DeeplabV3+ algorithm, this application uses expansion prediction for the processed images, which can effectively solve the "missegmentation" problem of large-size images and greatly improve the segmentation accuracy; a convolutional attention mechanism module CBAM is added to the decoding part to further strengthen the recognition of key features of broken grains and impurities, effectively improving problems such as misidentification and missed detection; a detail detection head is added to the framework to help the model better detect the edges, corners and other information of the segmented objects, refine the segmentation results, and at the same time enable the model to strengthen the learning of the overall contour features of the target object, which is conducive to improving the classification accuracy.

[0037] (2) The wheat image acquisition device designed in the present application adopts a sampling-discarding structure to realize the continuous acquisition of multiple samples of wheat images, ensuring that the wheat grains are evenly spread to reduce the influence of overlapping, occlusion and adhesion, and increase the accuracy of image detection.

[0038] (3) The online detection method of wheat impurity content and breakage rate provided in this application can provide a basis for the operator to intelligently adjust the operating parameters of the combine harvester, thereby improving the operating quality of the combine harvester and reducing losses and waste during wheat harvesting. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A schematic diagram of the structure of the wheat impurity and breakage detection system based on the improved DeeplabV3+ provided in this application;

[0040] Figure 2 A schematic diagram of the three-dimensional structure of the wheat image acquisition device provided in this application;

[0041] Figure 3 A front cross-sectional view of the wheat image acquisition device provided in this application;

[0042] Figure 4 A right side cross-sectional view of the wheat image acquisition device provided in this application;

[0043] Figure 5 The installation diagram of the wheat image acquisition device provided for this application;

[0044] Figure 6 The workflow diagram of the wheat image acquisition device provided in this application;

[0045] Figure 7 The workflow diagram of the wheat impurity-breakage detection method based on the improved DeeplabV3+ provided in this application;

[0046] Figure 8 Improved DeeplabV3+ algorithm workflow diagram provided for this application;

[0047] Fig. 9 Schematic diagram of the improved DeeplabV3+ algorithm network structure provided for this application;

[0048] Fig.10 The display device display interface provided by this application;

[0049] In the figure, 1- hopper; 2- servo; 3- diamond seat bearing; 4- rotating shaft; 5- conveying plate; 6- electromagnetic vibrator; 7- industrial camera; 8- COB light source; 9- camera bracket; 10- box top; 11- box; 12- discharge port; 13- grain outlet; 14- device bracket; 15- grain box. DETAILED DESCRIPTION

[0050] The present application is further described below in conjunction with the accompanying drawings and specific embodiments, but the protection scope of the present application is not limited thereto.

[0051] like Figure 1 As shown, a wheat impurity-breakage detection system based on improved DeeplabV3+ includes a wheat image acquisition device, an embedded processor and a display device. The wheat image acquisition device includes a wheat collection device, a wheat conveying device, an image acquisition device and a fixed component. The embedded processor is connected to the wheat collection device, the wheat conveying device, the image acquisition device and the display device respectively. The embedded processor is used to control the rotation angle of the servo of the wheat collection device to realize wheat feeding and control the vibration of the electromagnetic vibrator of the wheat conveying device to make the wheat flat within the field of view of the industrial camera. The embedded processor is also used to process the static wheat image transmitted by the industrial camera to obtain the impurity rate and breakage rate of the wheat, and finally send the wheat impurity rate and breakage rate to the display device for display; the embedded processor and the display device are installed in the cab.

[0052] like Figure 2-4As shown, the wheat collecting device of the present application includes a hopper 1, a steering gear 2, a diamond seat bearing 3 and a rotating shaft 4; the wheat conveying device includes a conveying plate 5 and an electromagnetic vibrator 6; the image acquisition device includes an industrial camera 7 and a COB light source 8; the fixed parts include a camera bracket 9, a box top 10, a box 11 and a discharge port 12;

[0053] like Figure 2 , Figure 3 , Figure 4 As shown, the upper left side of the box body 11 is provided with an opening and the right side is provided with a discharge port 12, the upper end of the box body 11 is connected to the lower end of the box body top 10, the box body top 10 is rectangular, and the upper end is provided with an opening; the hopper 1 is installed on the rotating shaft 4, the rotating shaft 4 is arranged at the upper left side opening of the box body 11, the two sides of the rotating shaft 4 pass through the opening of the box body 11 and are connected to the inside of the diamond seat bearing 3, the diamond seat bearing 3 is connected to the outside of the box body 11, the input shaft of the steering gear 2 is connected to the rear side of the rotating shaft 4, and the steering gear 4 rotates the hopper 1 filled with wheat to a certain angle and dumps it into the conveying plate 4 to complete the feeding; the conveying plate 4 is installed inside the box body 11, and the lower part of the conveying plate 4 The end is connected to an electromagnetic vibrator 6, and the lower end of the electromagnetic vibrator 6 is connected to the bottom of the box 11. The electromagnetic vibrator 6 makes the wheat grains evenly spread by adjusting parameters such as frequency and amplitude to reduce the influence of overlapping, occlusion and adhesion, and increase the accuracy of image detection; the industrial camera 7 is fixed above the top 10 of the box through the camera bracket 9, and the fixed position is in the vertical direction of the center of the conveying plate 4. The industrial camera 7 passes through the top 10 of the box and the lens faces the upper end of the conveying plate 4. The COB light source 8 is installed on the inner wall of the inclined surface of the top 10 of the box. The COB light source 8 is in the shape of a long strip, which can ensure sufficient light and uniform brightness during image acquisition.

[0054] like Figure 5 The figure shows the installation diagram of the wheat image acquisition device in the grain box 15. The box body 11 is fixed to the middle of the grain box 15 through the device bracket 14. The hopper 1 is located below the grain outlet 13 and is within the range of the wheat sprayed out of the grain outlet 13. Figure 6As shown, when the impurity content and breakage rate of wheat are detected online, wheat is ejected from the grain outlet 13 and falls into the hopper 1. After the hopper 1 is filled with wheat, the embedded processor controls the steering gear 2 to rotate, and the shaft 4 drives the hopper 1 to rotate a certain angle and then completely dumps the wheat grains onto the conveying plate 4. The embedded processor controls the electromagnetic vibrator 6 to generate an exciting force. While the wheat grains vibrate, they complete a linear motion on the conveying plate 4 to ensure that the wheat grains are evenly flattened and moved within the shooting range of the industrial camera 7. The industrial camera 7 collects a static wheat image under the illumination of the COB light source 8 and transmits it to the embedded processor. The embedded processor calls the improved DeeplabV3+ algorithm to process the static wheat image and obtain the impurity content and breakage rate of wheat. The impurity content, breakage rate and static wheat image are displayed on the display device. After the industrial camera 7 completes the static wheat image acquisition, under the continuous vibration of the electromagnetic vibrator 6, the wheat grains move quickly to the discharge port 12 on the conveying plate 4 to complete the sample abandonment operation.

[0055] The present application also provides a wheat impurity-breakage detection method based on improved DeeplabV3+, such as Figure 7 As shown, the main steps of this method are as follows:

[0056] S1, after the detection starts, the system is initialized first, including port initialization, graphics card hot start and thread pool initialization;

[0057] S2, after initialization, the interface is displayed;

[0058] S3, the embedded processor controls the steering gear to rotate and feed the wheat, dumps the wheat onto the conveying plate, and then controls the hopper to return to the initial position;

[0059] S4, the embedded processor sets the vibration time of the electromagnetic vibrator to make the wheat evenly spread and within the sampling field of view of the industrial camera;

[0060] S5, the industrial camera collects static wheat images and transmits them to the embedded processor;

[0061] S6, the embedded processor starts subthread 2 and calls the improved DeeplabV3+ algorithm to perform model inference on the static wheat image transmitted by the industrial camera, which mainly includes expansion prediction pre-processing, model inference and expansion prediction post-processing, such as Figure 8 As shown, the specific steps are as follows:

[0062] S601, performing padding, cropping and scaling operations on the static wheat image to output 15 512 pixel×512 pixel pre-processed wheat images and splicing them into multi-dimensional data and inputting them simultaneously;

[0063] S602, calling the improved DeeplabV3+ algorithm to perform model inference to achieve image segmentation of intact grains, broken grains, wheat husks, wheat straws, etc.;

[0064] S603, determining the output segmentation map of each pixel category, splicing the segmentation maps to the original size to obtain a final segmentation map, and then calculating the impurity rate and fragmentation rate according to the final segmentation map;

[0065] S7, when the embedded processor performs model inference, the main thread starts the electromagnetic vibrator to complete the sample discarding, waits for the execution of each thread to end, and outputs the final segmentation map and the calculated values ​​of the impurity rate and fragmentation rate.

[0066] S8, the embedded processor starts sub-thread 3 to process the output data, update the display device interface image display, numerical display, curve change and alarm display, and the main thread performs the next sampling operation.

[0067] like Fig. 9 The figure shows the improved DeeplabV3+ algorithm network structure. The original DeeplabV3+ algorithm does not adequately identify the key features of broken kernels and impurities, which easily leads to misidentification and missed detection. In addition, the algorithm lacks spatial details during segmentation, which causes adhesion in the segmentation results and confusion in the segmentation boundaries of different targets. Therefore, this application uses expansion prediction for the processed images based on the DeeplabV3+ algorithm to effectively solve the "missegmentation" problem of large-size images; adds a convolutional attention mechanism module CBAM to the decoding part to further strengthen the recognition of key features of broken kernels and impurities; adds a detail detection head part to the framework to help the model better detect the edges, corners and other information of the segmented objects, refine the segmentation results, and help improve the classification accuracy.

[0068] The improved DeeplabV3+ algorithm refers to the DeeplabV3+ algorithm as the baseline, introducing expansion prediction, convolutional attention mechanism module CBAM and detail detection head, improving the model's ability to learn key features of large-size images, making the segmentation results more refined, and achieving accurate recognition of broken kernels that are difficult to identify. The specific implementation is as follows:

[0069] Use dilation prediction for images processed by DeeplabV3+;

[0070] Add a convolutional attention mechanism module CBAM to the decoding part of DeeplabV3+;

[0071] Add detail detection head part to DeeplabV3+ framework.

[0072] Expansion prediction includes two parts: pre-processing and post-processing. The pre-processing of expansion prediction performs overlapping sliding window processing on the static wheat image to output 15 512 pixel × 512 pixel wheat pre-processed pictures and splices them into multi-dimensional data to input into the embedded processor. The post-processing of expansion prediction splices the central area of ​​each segmentation result image to obtain the pixel distribution of intact grains, broken grains, wheat husks, and wheat straws, and obtains the mass of each part according to the correspondence between the pixels and mass of intact grains, broken grains, wheat husks, and wheat straws, and calculates the impurity rate and breakage rate of wheat through the mass of each part.

[0073] The specific calculation formula for the impurity rate and breakage rate of wheat is:

[0074]

[0075]

[0076] Among them, R impurity is the impurity content, R broken is the crushing rate, m g is the mass of the whole kernel, m b is the mass of broken kernels, m s is the mass of the straw, m h For the quality of wheat husk.

[0077] The convolutional attention mechanism module CBAM includes a channel attention module and a spatial attention module; the channel attention module processes the deep feature map to obtain the channel attention weight and multiplies the channel attention weight with each channel of the deep feature map; the spatial attention module processes the deep feature map output by the channel attention module to obtain the spatial attention weight and multiplies it with the input feature map of the convolutional attention mechanism module CBAM, thereby optimizing the spatial information and channel information features of the deep feature map, reducing the interference of invalid features, and improving the utilization of key features.

[0078] The detail detection head is as follows: the shallow feature map is sequentially output through a 3×3 convolution layer, a BN layer, a ReLU activation layer, and a 1×1 convolution layer to output a detail feature map. At the same time, the Laplacian operator with a step size of 1, 2, and 4 is used to obtain the multi-scale detail information of the label map. The detail feature map is upsampled to the original size, and the predicted details are converted into a binary label edge map through a 1×1 convolution. Finally, the detail feature map and the binary label edge map are combined to calculate the detail loss value, so that more detail information can be learned during the upsampling process. The detail detection head only guides the direction of gradient descent in the training phase and plays a role in auxiliary training. The test phase does not include this part, so it does not increase the complexity of the model, and keeps the model inference time unchanged while improving the segmentation accuracy.

[0079] The specific calculation formula of the total loss function is:

[0080] Loss = F s +D s +F d +D d

[0081] Among them, Loss is the total loss function, F s The segmentation loss value calculated by Focal Loss, D s The segmentation loss value calculated by Dice Loss, F d The detail loss value calculated for Focal Loss, D d Detail loss value calculated for Dice Loss.

[0082] The improved DeeplabV3+ algorithm processes images as follows:

[0083] The static wheat image is cropped by dilation prediction pre-processing and reduced to 512 pixels × 512 pixels as model input. The backbone feature extraction uses an 8-fold down-sampling Xception-65 network to output shallow feature maps and deep feature maps. The deep feature map outputs channel weights and spatial weights through the CBAM module, and the weights are multiplied with the corresponding dimensions of the feature map to optimize the feature utilization in different dimensions. The multi-scale context information of the deep feature map is captured by the ASPP module to adapt to the segmentation targets of different sizes such as grains and impurities in the predicted image. The encoding part is completed after the multi-scale information is fused. In the decoding part, the shallow feature map output by the backbone network is introduced to provide more spatial location information for the upsampling of the deep feature map, and the detail features of the shallow feature map are captured by the detail detection head. The detail loss is calculated in combination with the binary label edge map. The loss value is added to the segmentation loss value to guide the feature map to learn more detail features during the upsampling process, thereby improving the segmentation accuracy of the model. After a single prediction is completed, the expansion prediction post-processing is performed to enlarge the predicted segmentation map to 1024 pixels × 1024 pixels, intercept the central area of ​​870 pixels × 870 pixels, and slide the cropping window to predict the remaining small-size images in turn until all image blocks of the static wheat image are completed. The intercepted central area is spliced ​​to output the final predicted segmentation result.

[0084] like Fig.10The display interface of the display device of the present application is shown. The upper left corner of the interface displays the model segmentation effect in real time. The model predicted segmentation image and the original image are transparently fused, and the generated fused image is easier to observe the accuracy of the model segmentation result. The lower left corner displays the numerical results of wheat impurity content and breakage rate and the detection warning results. When the impurity content or breakage rate exceeds the upper limit of the operation technical standard requirements, the warning state is red, otherwise it is green. The upper right corner is the change curve of the wheat impurity content and breakage rate data, which is used to record and save the change trend of the wheat impurity content and breakage rate data. When the change trend is abnormal, the operator should adjust the combine harvester operation parameters in time. At the bottom right are two "Start" and "End" buttons, which are used to start or stop the wheat impurity-breakage detection system.

[0085] The embodiments are preferred implementations of the present application, but the present application is not limited to the above implementations. Any obvious improvements, substitutions or modifications that can be made by those skilled in the art without departing from the essential content of the present application are within the scope of protection of the present application.

Claims

1. A wheat impurity-breakage detection method based on improved DeeplabV3+, characterized in that: The embedded processor calls the improved DeeplabV3+ algorithm to process the static wheat image transmitted by the industrial camera to obtain the impurity content and breakage rate of the wheat, which are displayed on the display device; Among them, the improved DeeplabV3+ algorithm processes the static wheat image transmitted by the industrial camera, including: the static wheat image is pre-processed by expansion prediction to output a 512 pixel × 512 pixel wheat pre-processed picture, and the wheat pre-processed picture is input into the backbone network Xception-65, and then a deep feature map and a shallow feature map are output; the deep feature map is processed by the convolutional attention mechanism module CBAM and the atrous spatial convolution pooling pyramid ASPP, and the deep feature map and the shallow feature map are decoded and processed, and a segmentation loss value is output; the shallow feature map is processed by the detail detection head to output a detail loss value; the total loss value is calculated by the total loss function, and finally the impurity rate and breakage rate of the wheat are calculated by the post-processing of the expansion prediction.

2. The wheat impurity-breakage detection method based on improved DeeplabV3+ according to claim 1, characterized in that: The improved DeeplabV3+ algorithm is used to achieve: The dilation prediction is applied to the images processed by DeeplabV3+; Add a convolutional attention mechanism module CBAM to the decoding part of DeeplabV3+; Add detail detection head part to DeeplabV3+ framework.

3. The wheat impurity-breakage detection method based on improved DeeplabV3+ according to claim 2, characterized in that: The expansion prediction includes two parts: the pre-processing and the post-processing; The pre-processing of the expansion prediction performs overlapping sliding window processing on the static wheat image, outputs 15 wheat pre-processing pictures of 512 pixels × 512 pixels, splices them into multi-dimensional data and inputs them into the embedded processor, and the embedded processor performs model inference operation on the multi-dimensional data to obtain each segmentation result map; The post-processing of the expansion prediction splices the central areas of the segmentation result images, and calculates the impurity rate and breakage rate of the wheat using a pixel quality mathematical model based on the pixel distribution of complete grains, broken grains, wheat husks, and wheat straws.

4. The wheat impurity-breakage detection method based on improved DeeplabV3+ according to claim 3, characterized in that: The method of using a pixel quality mathematical model to calculate the impurity rate and breakage rate of the wheat includes: The mass of each part is obtained through the correspondence between the pixels and the mass of the complete kernels, broken kernels, wheat husks, and wheat straws, and the impurity content and breakage rate of the wheat are calculated based on the mass of each part.

5. The wheat impurity-breakage detection method based on improved DeeplabV3+ according to claim 4, characterized in that: The calculation formula of the impurity content is: Among them, R impurity is the impurity content, m g is the mass of the whole kernel, m b is the mass of broken kernels, m s is the mass of the straw, m h For the quality of wheat husk.

6. The wheat impurity-breakage detection method based on improved DeeplabV3+ according to claim 4, characterized in that: The calculation formula of the crushing rate is: Among them, R broken is the crushing rate, m g is the mass of the whole kernel, m b The mass of broken kernels.

7. The wheat impurity-breakage detection method based on improved DeeplabV3+ according to claim 2, characterized in that: The convolutional attention mechanism module CBAM includes a channel attention module and a spatial attention module; The channel attention module is used to process the deep feature map, obtain the channel attention weight and multiply the channel attention weight with each channel of the deep feature map; The spatial attention module processes the deep feature map output by the channel attention module to obtain the spatial attention weight and multiplies the spatial attention weight with the input feature map of the convolutional attention mechanism module CBAM.

8. The wheat impurity-breakage detection method based on improved DeeplabV3+ according to claim 2, characterized in that: The detail detection head part is used to implement: The shallow feature map sequentially outputs a detail feature map through a 3×3 convolutional layer, a BN layer, a ReLU activation layer, and a 1×1 convolutional layer. At the same time, a Laplacian operator with a step size of 1, 2, and 4 is used to obtain multi-scale detail information of the label map, and the detail feature map is upsampled to the original size. The predicted details are converted into a binary label edge map through a 1×1 convolution. Finally, the detail feature map and the binary label edge map are combined to calculate the detail loss value.

9. The wheat impurity-breakage detection method based on improved DeeplabV3+ according to claim 1, characterized in that: The calculation formula of the total loss function is: Loss=F s +D s +F d +D d ; Among them, Loss is the total loss function, F s is the segmentation loss value calculated by Focal Loss, D s The segmentation loss value calculated by Dice Loss, F d The detail loss value calculated for the Focal Loss, D d The detail loss value calculated for the Dice Loss.

10. A wheat impurity and breakage detection system based on improved DeeplabV3+, characterized in that: It comprises a wheat image acquisition device, an embedded processor and a display device, wherein the embedded processor is connected to the wheat image acquisition device and the display device respectively, and the embedded processor is used to process the static wheat image transmitted by the industrial camera using the improved DeeplabV3+ algorithm to obtain the impurity content and breakage rate of the wheat; The wheat impurity and breakage detection system based on improved DeeplabV3+ is used to implement the wheat impurity and breakage detection method based on improved DeeplabV3+ as described in any one of claims 1 to 9.

Images