Crop row real-time detection method based on space and context information deep fusion image segmentation

By adopting an image segmentation method based on deep fusion of spatial and context information in crop row detection, combining the dual-branch structure and convolutional neural network, the problem of slow detection speed in the existing technology is solved, and the Accurate and accurate inspection, suitable for agricultural machinery.

CN119007010AActive Publication Date: 2024-11-22BEIJING INST OF TECH

Patent Information

Application Number
CN202411076128.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-07
Publication Date
2024-11-22
Estimated Expiration
2044-08-07

AI Technical Summary

Technical Problem

The existing crop row detection methods are difficult to balance between accuracy and real-time. Especially for different crops, different growth periods and different environments, the existing methods have a slow detection speed and cannot be implemented in real-time deployment on agricultural machinery. .

Method used

Using an image segmentation method based on deep fusion of spatial and context information, the lightweight dual-branch structure is designed, combined with convolutional neural network and channel attention mechanism, the fusion of spatial and context information is achieved, and useless information interference is eliminated and detection is improved. Accuracy and real-time.

Benefits of technology

It achieves the accuracy of crop inspections for different crops, different growth periods and different environments, while ensuring the real-time detection, improving the detection speed, and suitable for the deployment of agricultural machinery.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119007010A_ABST
    Figure CN119007010A_ABST
Patent Text Reader

Abstract

The invention relates to the field of image processing, in particular to a crop row real-time detection method based on space and context information deep fusion image segmentation. The extraction of space and context information is realized by designing a double-branch structure. Wherein the low-resolution branch comprises a context extractor, and multi-scale information is fused in parallel to realize efficient extraction of long-range semantics. An information fusion module is constructed by utilizing channel information with obvious image features in each growth cycle of crops, deep ordered fusion of double branches is carried out, useless information interference is removed, crop row detection for different crops in different periods and different environments is realized, and the accuracy is realized while the real-time performance is ensured.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a real-time crop row detection method based on image segmentation with deep fusion of spatial and contextual information. Background Art

[0002] With the development of smart agriculture, crop row detection has become an important part of agricultural intelligent equipment. It is widely used in agricultural production such as assisted navigation, precision fertilization, weeding, spraying, and harvesting. There are three main methods for crop row detection at present, namely, based on traditional image processing, based on target detection, and based on image segmentation. The method based on traditional image processing is limited by manually set rules, such as converting the image into HSV color space and manually selecting the threshold to extract green crops, which is easily interfered by factors such as lighting, shooting angle and weeds, and it is difficult to meet the accuracy requirements; the method based on target detection requires complex stringing post-processing, the algorithm is time-consuming, and it is not suitable for crop row recognition with small spacing due to the late growth period. The method based on image segmentation is suitable for crop row detection in different growth periods, and can effectively solve problems such as complex lighting and weeds, so it is suitable for crop row detection.

[0003] Existing crop row detection methods based on image segmentation usually directly adopt general image segmentation models, such as

[0004] The Unet model is used in "Automated detection of crop-row lines and measurement of maize width for boom spraying", which has a large number of parameters and is only suitable for offline detection; the Deeplabv3+ model is used in "Fusing vegetation index and ridge segmentation for robust vision based autonomous navigation of agricultural robots in vegetable farms", which is a network designed specifically for the segmentation of large-resolution images. It uses dilated convolutional layers and has a significant effect on the segmentation of large objects, but is not suitable for the segmentation of small crop rows. The existing crop row segmentation method does not design the network structure according to the characteristics of the crop scene, resulting in high detection accuracy but slow detection speed, and the algorithm cannot be deployed on agricultural machinery.

[0005] Therefore, a crop row detection method that can be used for different crops, different periods, and different environments, and this method can achieve accuracy while ensuring real-time performance, has become an urgent need in this field. Summary of the invention

[0006] Based on the above purpose, the purpose of the present invention is to overcome the shortcomings existing in the prior art, and propose a real-time crop row detection method based on image segmentation with deep fusion of spatial and contextual information, and realize the extraction of spatial and contextual information by designing a lightweight dual-branch structure. The low-resolution branch includes a context extractor, and the multi-scale information is fused in parallel to realize the efficient extraction of long-range semantics. An information fusion module is constructed by utilizing the channel information with obvious characteristics of the images of each growth cycle of crops, and a deep and orderly fusion of the two branches is performed to remove useless information interference, thereby realizing crop row detection for different crops, different periods, and different environments, achieving accuracy while ensuring real-time performance.

[0007] To this end, the first technical solution of the present application discloses a real-time crop line detection method based on deep fusion image segmentation of spatial and contextual information, comprising:

[0008] S1. Collect target image;

[0009] S2. Based on convolutional neural network, deep fusion of spatial and contextual information of target image features;

[0010] S3. Use the segmentation head to perform pixel-level segmentation on the fused image features, and perform line fitting on the segmentation results to obtain crop lines.

[0011] Furthermore, the acquisition in S2 performs deep fusion of spatial and contextual information of the target image as follows:

[0012] (1) Differentiate the target image features into a spatial information branch and a contextual information branch through a convolutional neural network;

[0013] (2) Deeply integrate the spatial information and contextual information of the two branches;

[0014] (3) Combine the fused spatial information and contextual information to obtain the fused image features.

[0015] Furthermore, the step (1) is: performing a first convolutional layer process of downsampling the target image by a factor of 2, and then further extracting the target image features through a second convolutional layer process, and further differentiating the network into a spatial information branch and a contextual information branch.

[0016] Furthermore, the step (2) is: processing the context information through the third convolution layer to achieve further information extraction, and downsampling the spatial information so that the spatial information feature is consistent with the context information feature size, at this time the spatial information is fused with the context information as a guiding role, and further extracting the first context fusion information through the fourth convolution layer; upsampling the first context fusion information so that the first context information feature is consistent with the spatial information feature size, at this time the first context information is fused with the spatial information as a guiding role, and further extracting the first spatial fusion information through the fifth convolution layer; repeating the above steps, downsampling the first spatial fusion information, fusing it with the first context information as a guiding role, and extracting the second context fusion information through the sixth convolution layer; upsampling the second context fusion information, fusing it with the first spatial information as a guiding role, and extracting the second spatial fusion information through the seventh convolution layer; aggregating the obtained second context fusion information to obtain context aggregation information, thereby obtaining an efficient global representation of context information;

[0017] The context aggregation information is combined with the second spatial fusion information to obtain fused image features.

[0018] Furthermore, the fusion is to establish different weight coefficients between channels for spatial information and contextual information through a channel attention mechanism, and then generate weights through sigmoid through a gating unit, and obtain the output after weighting the two parts of information.

[0019] Furthermore, the aggregation operation is: the second context fusion information is processed by the context information aggregation module to obtain context aggregation information; the specific operation is: the second context fusion information is input with a feature image of 1 / 32 image resolution, and the pooling operations of ① pooling kernel = 5, step size = 2, ② pooling kernel = 9, step size = 4, ③ pooling kernel = input feature resolution size are performed in parallel to generate features of 1 / 64, 1 / 128, and 1 / 256 image resolutions, and further subjected to 1x1 convolution operations respectively, and the processed features are added in parallel with the features obtained after the initial features are subjected to 1x1 convolution operations. The four features after parallel addition are subjected to 3x3 convolution operations respectively and connected to achieve rapid aggregation of context information over a long distance.

[0020] Furthermore, the line fitting in S3 is as follows: the segmented image is converted into a bird's-eye view through external parameters, the bird's-eye view is vertically cumulatively projected to obtain the number of crop rows, and then the lane line points are searched with the peak point as the starting point through the sliding window method, and further the non-zero pixels in each sliding window are fitted using the least squares method to obtain the crop row lines.

[0021] The second technical solution of the present application discloses a crop row detection system based on image segmentation with deep fusion of spatial and contextual information, which is used to implement the above-mentioned crop row detection method based on image segmentation with deep fusion of spatial and contextual information, including:

[0022] Image acquisition module: used to acquire target images;

[0023] Convolutional neural network module: including the first to seventh convolutional neural network models, used for extracting features of the target image or the differentiated spatial information and context information; and

[0024] Fusion module: including spatial information fusion module and context information fusion module, as well as channel attention mechanism module and gate unit module for allocating weights during fusion, to achieve effective fusion of spatial information and context information;

[0025] Context aggregation module: aggregates the obtained second context fusion information to obtain an efficient global representation of the context information.

[0026] Image segmentation module: used to perform pixel-level classification prediction on the fused image features.

[0027] The third technical solution of the present application discloses a computer device, comprising:

[0028] Memory; and

[0029] A processor is configured to execute the above-mentioned crop row real-time detection method based on spatial and context information deep fusion image segmentation based on instructions stored in the memory.

[0030] The fourth technical solution of the present application discloses a computer-readable storage medium, which stores computer instructions. When the instructions are executed by a processor, the above-mentioned real-time crop row detection method based on deep fusion image segmentation of spatial and contextual information can be implemented.

[0031] Beneficial effects: This application realizes the extraction of spatial and contextual information by designing a lightweight dual-branch structure. The low-resolution branch contains a context extractor, which fuses multi-scale information in parallel to achieve efficient extraction of long-range semantics. Using the channel information with obvious characteristics of images of crops in each growth cycle, an information fusion module is constructed to perform deep and orderly fusion of the two branches to remove useless information interference, thereby realizing crop row detection for different crops, different periods, and different environments, achieving accuracy while ensuring real-time performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A flow chart of the steps of a crop row detection method;

[0033] Figure 2 It is the overall framework of a crop row detection system based on image segmentation with deep fusion of spatial and contextual information;

[0034] Figure 3 It is a schematic diagram of the fusion module;

[0035] Figure 4 It is a schematic diagram of context information aggregation module;

[0036] Figure 5 is a specific input image and the related result image. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0038] Unless otherwise specifically stated, the relative arrangement, numerical expressions and numerical values ​​of the parts and steps set forth in these embodiments do not limit the scope of the present application. At the same time, it should be understood that, for ease of description, the sizes of the various parts shown in the accompanying drawings are not drawn according to the actual proportional relationship. The technology, method and equipment known to those of ordinary skill in the relevant field may not be discussed in detail, but in appropriate cases, the technology, method and equipment should be considered as a part of the authorization specification. In all examples shown and discussed here, any specific value should be interpreted as being merely exemplary, rather than as a limitation. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters represent similar items in the following drawings, so that once a certain item is defined in an accompanying drawing, it does not need to be further discussed in subsequent drawings.

[0039] Unless otherwise specified, the meanings of the technical terms in this specification are the same as those generally understood by those skilled in the art. However, if there is any conflict, the definitions in this specification shall prevail.

[0040] To this end, the first technical solution of the present application discloses a real-time crop line detection method based on deep fusion image segmentation of spatial and contextual information, such as Figure 1 As shown, including:

[0041] S1. Collect target image;

[0042] S2. Based on convolutional neural network, deep fusion of spatial and contextual information of target image is performed;

[0043] S3. Input the fused image features into the segmentation head to realize image segmentation, and perform line fitting on the segmentation results to obtain crop row lines.

[0044] In the implementation mode of the present application, the purpose of collecting the target image is to construct a crop row data set to provide sufficient samples for the later training model. In some preferred embodiments, the image acquisition method is generally achieved by using the DJI Osmo Mobile 6 gimbal, setting different pitch angles, and setting different acquisition heights to simulate the data collection state of different types of agricultural robots, and enhancing the generalization performance of the algorithm after training with this data set, so that the algorithm can be directly deployed in various agricultural robots to perform row perception tasks. In addition, an ordinary mobile phone is used for image capture, with an image resolution of 1920×1080 and a frame rate of 30Fps, which is consistent with the image output format of the common cheap RGB cameras on the market. The farmland row perception data set covers a variety of scene types, including different shooting angles, different crop types, different lighting conditions, different growth conditions, and other special cases. For the shooting angle, we collect data at different pitch angles and yaw angles. As the pitch angle becomes smaller and smaller, the number of rows involved becomes smaller and smaller, covering the row detection needs of different agricultural robots; for crop types, we collected a variety of common crops, including wheat, corn, eggplant, pepper, potato, soybean, and cabbage; for light conditions, it is divided into day and night. There are different light conditions of different strengths at different times of the day, and at night, searchlights are used to simulate the light conditions of night operations. For growth conditions, we collected wheat and corn throughout their life cycle. For other crops, typical growth periods were collected. For other situations, including whether there is strong wind, whether the surface is wet, crop row interruptions, whether there are weeds, and whether there are other artificial visual recognition interferences. The collected images are marked with multi-line segments by annotation tools, and the lines are given different widths according to the growth status of the crops, and converted into binary mask images.

[0045] In some specific implementations, the acquisition in S2 performs deep fusion of spatial and contextual information of the target image as follows:

[0046] (1) Differentiate the target image into a spatial information branch and a contextual information branch through a convolutional neural network;

[0047] (2) Deeply integrate the spatial information and contextual information of the two branches;

[0048] (3) Combine the fused spatial information and contextual information to obtain the fused image features.

[0049] In this embodiment, the step (1) is: the target image is processed by the first convolution layer with 2 times downsampling, and then further extracted from the target image by the second convolution layer, and the network is further differentiated into a spatial information branch and a context information branch.

[0050] In this embodiment, the first convolution layer is composed of 3x3 convolutions to achieve 2x downsampling, and the second convolution layer converts the feature size to 1 / 4 of the original image size, and further differentiates the network into a spatial information branch and a context information branch. Figure 2 As shown in the figure, the target image is differentiated into a spatial information branch and a context information branch after the convolution layer. In order to maintain a higher resolution to retain the spatial information, the convolution step is set to 1 to keep the output size of the spatial information unchanged, thereby maintaining a higher resolution. For the context information branch, the convolution step is set to 2 to make the image features smaller and smaller, so as to further extract the context information.

[0051] Further, such as Figure 2 As shown, the step (2) is: processing the context information through the third convolution layer to achieve further information extraction, and downsampling the spatial information so that the spatial information feature is consistent with the context information feature size, at this time the spatial information is fused with the context information as a guiding role, and further extracting the first context fusion information through the fourth convolution layer; upsampling the first context fusion information so that the context information feature is consistent with the spatial information feature size, at this time the context information is fused with the spatial information as a guiding role, and further extracting the first spatial fusion information through the fifth convolution layer; repeating the above steps, downsampling the first spatial fusion information, fusing it with the context information as a guiding role, and extracting the second context fusion information through the sixth convolution layer; upsampling the second context fusion information, fusing it with the spatial information as a guiding role, and extracting the second spatial fusion information through the seventh convolution layer;

[0052] In this embodiment, the reason why spatial information is first used as a guiding role to fuse with context information is that in the early stage of differentiation, the feature resolution of context information is not low enough, and the extracted high-level semantic information is not sufficient, while spatial information maintains a higher resolution and retains relatively complete spatial information, so the spatial information is unidirectionally input into the context information to better capture the context information. After the spatial information and the context information are fused for the first time and the first context fusion information is obtained through the convolution layer, its feature size is processed to 1 / 16 of the image size. At this time, the context information obtains certain semantic information and can provide certain semantic information for the spatial information branch. Then, the context information is unidirectionally input into the spatial branch to obtain the first spatial fusion information to better capture the spatial information. Then, the above steps are repeated to obtain the second context fusion information and the second spatial fusion information, so that the spatial and context information of the image can be fully mined.

[0053] Furthermore, the fusion is to establish different weight coefficients between channels for spatial information and contextual information through a channel attention mechanism, and then generate weights through sigmoid through a gating unit, and obtain the output after weighting the two parts of information.

[0054] It is understandable that the reason for giving weights to different channels during fusion is that crops have consistent color characteristics at different growth stages. For example, different types of crops are nearly green at the beginning of growth. Therefore, in the fusion stage, the two information features are firstly used to establish the dependency between channels using the channel attention mechanism, such as Figure 3 As shown in the figure: Different weight coefficients are assigned to each channel, so as to effectively strengthen important features and suppress unimportant features. Furthermore, a gating unit fusion mechanism is designed to generate weights for the two inputs through sigmoid, and the weighted sum of the two parts is obtained to obtain the output. The resulting context output integrates spatial information and can better capture context information in the downstream stage. It should be noted that the channel attention mechanism of the fusion step of the whole process is consistent with the gating unit.

[0055] In a further implementation, the obtained second context fusion information needs to be aggregated to obtain an efficient global representation of the context information. Figure 4As shown in the figure, the second context fusion information is processed by the context information aggregation module to obtain context aggregation information; the specific operation is: the second context fusion information is input with a feature image of 1 / 32 image resolution, and the pooling operations of ① pooling kernel = 5, step size = 2, ② pooling kernel = 9, step size = 4, and ③ pooling kernel = input feature resolution size are performed in parallel to generate features of 1 / 64, 1 / 128, and 1 / 256 image resolutions, and further subjected to 1x1 convolution operations respectively, and the processed features and the features obtained after the initial features are subjected to 1x1 convolution operations are added in parallel. . The four features after parallel addition are subjected to 3x3 convolution operations respectively and connected to achieve rapid aggregation of context information over a long distance to obtain efficient global representation.

[0056] In a further implementation, the context aggregation information is merged with the second spatial fusion information to obtain fused image features.

[0057] In some specific embodiments, the image segmentation in S3 is performed by inputting the fused image features into an image segmentation head. In particular, during the model training process, an additional auxiliary segmentation head is added before the fifth convolutional layer to obtain an auxiliary cross entropy loss to achieve the optimization of the deep convolutional neural network. 辅助 The generated auxiliary cross loss is assigned a weight α and added to the cross entropy loss obtained by the normal segmentation head to obtain the final loss function, which can be expressed as L 最终 =L 常规 + L辅助 The auxiliary segmentation head does not appear in the test phase, but only in the training phase. The loss function is used to update the model parameters to obtain the final segmentation model.

[0058] In some specific embodiments, the line fitting in S3 is as follows: the segmented image is converted into a bird's-eye view through an external reference lane line segmentation map, the bird's-eye view map is vertically cumulatively projected to obtain the number of crop rows, and then the lane line points are searched with the peak point as the starting point through a sliding window method, and further the non-zero pixels in each sliding window are fitted using the least squares method to obtain the crop row lines.

[0059] It should be noted that the second to seventh convolutional layers of the present application are all composed of blueprint separation convolutional units, and the number of blueprint separation convolutional units, the number of output channels and the step size contained in each spacing neural network can be determined according to requirements.

[0060] The second embodiment of the present application discloses a crop row detection system based on image segmentation with deep fusion of spatial and contextual information, which is used to implement the crop row detection method based on image segmentation with deep fusion of spatial and contextual information, including:

[0061] Image acquisition module: used to acquire target images;

[0062] Convolutional neural network module: including the first to seventh convolutional neural network models, used for sampling the target image or the differentiated spatial information and context information image; and

[0063] Fusion module: including spatial information fusion module and context information fusion module, as well as channel attention mechanism module and gate unit module for allocating weights during fusion, to achieve effective fusion of spatial information and context information;

[0064] Context aggregation module: aggregates the obtained second context fusion information to obtain an efficient global representation of the context information.

[0065] Image segmentation module: used to perform pixel-level classification prediction on the fused image features.

[0066] The present invention is specifically described below by way of examples.

[0067] Figure 5 The first row in the figure is a picture of corn in a certain area of ​​Beijing. The shooting angle, height, light intensity and crop growth period of the picture are different, which reflects the richness and challenge of the dataset; the second row is the corresponding annotated picture. The crop rows are marked with multiple line segments through the annotation tool, and the lines are given different widths according to the growth status of the crops, and converted into a binary mask map; the third row is the result obtained after segmentation by the model trained by the method, which is compared with the annotation result in the second row, reflecting a better segmentation result; the fourth row is the bird's-eye view result after the segmentation map is transformed by the external parameter perspective; the fifth row is the result after the line fitting inverse perspective transformation. It can be seen that the fitting result is well aligned with the center of the crop row.

[0068] The following table shows the quantitative test results conducted on the self-collected data set. It can be seen that compared with Unet and Fast-SCNN, this algorithm has the fastest computing speed while maintaining high accuracy, achieving real-time detection of crop rows.

[0069] Segmentation method mIOU mAcc mRecall FPS Unet 75.32 75.10 73.45 15 Fast-SCNN 65.73 70.48 70.21 55 Algorithm of the present invention 73.49 76.45 75.16 65

[0070] On the other hand, the present application also discloses a computer device, comprising:

[0071] Memory; and

[0072] A processor is configured to execute the above-mentioned crop row real-time detection method based on spatial and context information deep fusion image segmentation based on instructions stored in the memory.

[0073] Finally, the present application also discloses a computer-readable storage medium, which stores computer instructions. When the instructions are executed by a processor, the above-mentioned real-time crop row detection method based on deep fusion image segmentation of spatial and contextual information can be implemented.

[0074] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A real-time crop line detection method based on spatial and context information deep fusion image segmentation, comprising: S1. Collect target image; S2. Based on convolutional neural network, deep fusion of spatial and contextual information of target image is performed; S3. Use the segmentation head to perform pixel-level classification prediction on the fused image features, and perform line fitting on the segmentation results to obtain crop row lines.

2. The method according to claim 1, characterized in that: S2 collects and deeply integrates the spatial and contextual information of the target image, specifically: (1) Differentiate the target image into a spatial information branch and a contextual information branch through a convolutional neural network; (2) Deeply integrate the spatial information and contextual information of the two branches; (3) Combine the fused spatial information and contextual information to obtain the fused image features.

3. The method according to claim 2, characterized in that: The step (1) is: subjecting the target image to a first convolutional layer processing with a 2-fold downsampling, and then subjecting the target image to a second convolutional layer processing for further feature extraction, and further differentiating the network into a spatial information branch and a contextual information branch.

4. The method according to claim 2, characterized in that: The step (2) is: processing the context information through the third convolutional layer to achieve further information extraction, and downsampling the spatial information so that the spatial information feature is consistent with the context information feature size, at this time the spatial information is fused with the context information as a guiding role, and further extracting the first context fusion information through the fourth convolutional layer; upsampling the first context fusion information so that the first context information feature is consistent with the spatial information feature size, at this time the first context information is fused with the spatial information as a guiding role, and further extracting the first spatial fusion information through the fifth convolutional layer; Repeat the above steps, downsample the first spatial fusion information, fuse it with the first context information as a guide, and extract the information through the sixth convolutional layer to obtain the second context fusion information; After upsampling the second context fusion information, it is fused with the first spatial information as a guide, and the second spatial fusion information is obtained by extracting information through the seventh convolutional layer; Aggregating the obtained second context fusion information to obtain context aggregation information, thereby obtaining an efficient global representation of the context information; The context aggregation information is combined with the second spatial fusion information to obtain fused image features.

5. The method according to claim 4, characterized in that: The fusion is to establish different weight coefficients between channels for spatial information and context information through the channel attention mechanism, and then generate weights through sigmoid through the gating unit, and obtain the output after weighting the two parts of information.

6. The method according to claim 4, characterized in that: The context aggregation operation is as follows: the second context fusion information is input with a feature image of 1 / 32 image resolution, and the pooling operations of ① pooling kernel = 5, step size = 2, ② pooling kernel = 9, step size = 4, and ③ pooling kernel = input feature resolution size are performed in parallel to generate features of 1 / 64, 1 / 128, and 1 / 256 image resolutions, and further subjected to 1x1 convolution operations respectively, and the processed features are added in parallel with the features obtained after the initial features are subjected to 1x1 convolution operations. The four features after parallel addition are subjected to 3x3 convolution operations respectively and connected to achieve rapid aggregation of context information over a long distance.

7. The method according to claim 1, characterized in that: The line fitting described in S3 is as follows: the segmented image is converted into a bird's-eye view through external parameters, the bird's-eye view is vertically cumulatively projected to obtain the number of crop rows, and then the lane line points are searched with the peak point as the starting point through the sliding window method, and the non-zero pixels in each sliding window are further fitted using the least squares method to obtain the crop row lines.

8. A crop row detection system based on spatial and context information deep fusion image segmentation, which is used to implement the method according to any one of claims 1 to 7, comprising: Image acquisition module: used to acquire target images; Convolutional neural network module: including the first to seventh convolutional neural network models, used for extracting features of the target image or the differentiated spatial information and context information image; and Fusion module: including spatial information fusion module and context information fusion module, as well as channel attention mechanism module and gate unit module for allocating weights during fusion, to achieve effective fusion of spatial information and context information; Context aggregation module: aggregates the obtained second context fusion information to obtain an efficient global representation of the context information. Image segmentation module: used to perform pixel-level classification prediction on the fused image features.

9. A computer device, characterized in that: include: Memory; as well as A processor configured to execute, based on instructions stored in the memory, a method for real-time detection of crop rows based on image segmentation with deep fusion of spatial and contextual information as described in any one of claims 1-7.

10. A computer-readable storage medium storing computer instructions, which, when executed by a processor, can implement the real-time crop row detection method based on image segmentation with deep fusion of spatial and contextual information as described in any one of claims 1-7.

Citation Information

Patent Citations

  • Context pyramid fusion network and image segmentation method

    CN110689083A

  • Crop row detection method and device based on deep learning image segmentation

    CN113128576A

  • Method for providing control data, a crop failure map and / or a replanting map

    WO2024133848A1

Cited By

  • Wheat infected wheat ear segmentation method based on perception enhanced gating network

    CN119339088A