A real-time crop row detection method based on deep fusion of spatial and contextual information for image segmentation.

By incorporating spatial and contextual information through a lightweight dual-branch structure design, the problem of balancing accuracy and real-time performance in existing crop row detection methods is solved, enabling efficient crop row detection under different conditions and making it suitable for agricultural machinery.

CN119007010BActive Publication Date: 2026-11-10BEIJING INST OF TECH
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202411076128.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-07
Publication Date
2026-11-10
Estimated Expiration
2044-08-07

AI Technical Summary

Technical Problem

Existing crop row detection methods struggle to balance detection accuracy and real-time performance. In particular, the current network architecture is inadequate for different crops, growth stages, and environments, resulting in slow detection speeds and making it impossible to deploy the algorithm on agricultural machinery.

Method used

A lightweight dual-branch structure is designed. By incorporating a context extractor in the low-resolution branch, multi-scale information is fused to construct an information fusion module. This module performs deep and ordered fusion of the two branches, removes useless information interference, and achieves efficient extraction of spatial and contextual information.

Benefits of technology

It enables accurate crop row detection for different crops, at different times, and under different environments, ensuring real-time detection and accuracy, and is suitable for deployment on agricultural machinery.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119007010B_ABST
    Figure CN119007010B_ABST
Patent Text Reader

Abstract

This invention relates to the field of image processing, specifically to a real-time crop row detection method based on deep fusion of spatial and contextual information in image segmentation. A dual-branch structure is designed to extract spatial and contextual information. The low-resolution branch contains a context extractor, which fuses multi-scale information in parallel to achieve efficient extraction of long-range semantics. Utilizing the distinct channel information of images from different crop growth stages, an information fusion module is constructed to perform deep, ordered fusion of the dual branches, removing irrelevant information interference. This enables crop row detection for different crops, different growth stages, and different environments, achieving both accuracy and real-time performance.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of image processing, and more specifically to a real-time crop row detection method based on deep fusion of spatial and contextual information for image segmentation. Background Technology

[0002] With the development of smart agriculture, crop row detection has become an important component of intelligent agricultural equipment. It is widely used in agricultural production, including assisted navigation, precision fertilization, weeding, spraying, and harvesting. Currently, there are three main methods for crop row detection: those based on traditional image processing, those based on object detection, and those based on image segmentation. Methods based on traditional image processing are limited by manually set rules, such as converting the image to the HSV color space and manually selecting thresholds to extract green crops. These methods are easily affected by factors such as lighting, shooting angle, and weeds, making it difficult to meet accuracy requirements. Methods based on object detection require complex post-processing, have high algorithmic time consumption, and are unsuitable for recognizing crop rows with small spacing due to late-stage growth. Image segmentation methods are adaptable to crop row detection at different growth stages and can effectively solve problems related to complex lighting and weeds, making them suitable for crop row detection.

[0003] Existing crop row detection methods based on image segmentation typically employ general image segmentation models, such as...

[0004] The paper "Automated detection of Crop-Row lines and measurement of maize width for boom spraying" uses the Unet model, which has a large number of parameters and is only suitable for offline detection. The paper "Fusing vegetationindex and ridge segmentation for robust vision-based autonomous navigation of agricultural robots in vegetable farms" uses the Deeplabv3+ model, a network specifically designed for high-resolution image segmentation. It uses dilated convolutional layers and is effective for segmenting large objects, but not for segmenting small crop rows. Existing crop row segmentation methods do not design their network structures according to the characteristics of the crop scene, resulting in high detection accuracy but slow detection speed, making it impossible to deploy the algorithms on agricultural machinery.

[0005] Therefore, a crop row detection method that can be used for different crops, different periods, and different environments, and that can ensure real-time performance while achieving accuracy, is urgently needed in this field. Summary of the Invention

[0006] Based on the above objectives, the purpose of this invention is to overcome the shortcomings of existing technologies and propose a real-time crop row detection method based on deep fusion of spatial and contextual information in image segmentation. This method achieves the extraction of spatial and contextual information through a lightweight dual-branch structure. The low-resolution branch contains a context extractor, and multi-scale information is fused in parallel to achieve efficient extraction of long-range semantics. Utilizing the distinct channel information of images from different crop growth stages, an information fusion module is constructed to perform deep, ordered fusion of the dual branches, removing useless information interference. This enables crop row detection for different crops, different growth stages, and different environments, achieving both accuracy and real-time performance.

[0007] Therefore, the first technical solution of this application discloses a real-time crop row detection method based on deep fusion of spatial and contextual information image segmentation, including:

[0008] S1. Acquire the target image;

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

[0010] S3. The fused image features are segmented at the pixel level using a segmentation head, and the segmented results are fitted with lines to obtain crop rows.

[0011] Furthermore, the acquisition described in S2, which involves deep fusion of spatial and contextual information of the target image, specifically includes:

[0012] (1) The target image features are differentiated into spatial information branches and contextual information branches through a convolutional neural network;

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

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

[0015] Furthermore, step (1) involves processing the target image with a first convolutional layer that downsamples by 2 times, then processing it with a second convolutional layer to further extract the target image features, and further differentiating the network into a spatial information branch and a contextual information branch.

[0016] Further, step (2) involves: processing the context information through a third convolutional layer to achieve further information extraction; downsampling the spatial information so that the spatial information features are the same size as the context information features; then, the spatial information acts as a guide and is fused with the context information; and further, the information is extracted through a fourth convolutional layer to obtain the first context fusion information; upsampling the first context fusion information so that the first context information features are the same size as the spatial information features; then, the first context information acts as a guide and is fused with the spatial information; and further, the information is extracted through a fifth convolutional layer to obtain the first spatial fusion information; repeating the above steps, downsampling the first spatial fusion information and fusing it with the first context information as a guide, and extracting the information through a sixth convolutional layer to obtain the second context fusion information; upsampling the second context fusion information and fusing it with the first spatial information as a guide, and extracting the information through a seventh convolutional layer to obtain the second spatial fusion information; and aggregating the obtained second context fusion information to obtain context aggregation information, thereby obtaining an efficient global representation of the context information;

[0017] The context aggregation information and the second spatial fusion information are merged to obtain the fused image features.

[0018] Furthermore, the fusion involves establishing different weight coefficients between channels by using a channel attention mechanism for spatial information and contextual information, then generating weights through a gating unit using a sigmoid function, and finally weighting the two parts of information to obtain the output.

[0019] Furthermore, the aggregation operation is as follows: the second context fusion information is processed by the context information aggregation module to obtain context aggregation information; specifically, the second context fusion information is used as input with a feature image at 1 / 32 image resolution, and pooling operations are performed in parallel with ① pooling kernel = 5, stride = 2, ② pooling kernel = 9, stride = 4, and ③ pooling kernel = the input feature resolution, generating features at 1 / 64, 1 / 128, and 1 / 256 image resolutions. These features are then further processed by 1x1 convolution operations, and the processed features are added in parallel with the features obtained after the initial features have undergone 1x1 convolution operations. The four features after parallel addition are then convolved with 3x3 convolution operations to achieve rapid aggregation of context information over a long distance.

[0020] Furthermore, the line fitting described in S3 is as follows: the segmented image is converted into a bird's-eye view through extrinsic parameters, the number of crop rows is obtained by vertically accumulating and projecting the bird's-eye view, and then the points of the lane line are searched using the sliding window method with the peak point as the starting point. Furthermore, the crop row line is obtained by fitting the non-zero pixels in each sliding window using the least squares method.

[0021] The second technical solution of this application discloses a crop row detection system based on deep fusion of spatial and contextual information for image segmentation. This system is used to implement the aforementioned crop row detection method based on deep fusion of spatial and contextual information for image segmentation, and includes:

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

[0023] Convolutional Neural Network Module: Includes convolutional neural network models one through seven, used for feature extraction from the target image or differentiated spatial and contextual information; and

[0024] The fusion module includes a spatial information fusion module, a contextual information fusion module, a channel attention mechanism module for weight allocation during fusion, and a gating unit module, which are used to achieve effective fusion of spatial and contextual 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 and prediction of the fused image features.

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

[0028] Memory; and

[0029] The processor is configured to execute the above-described real-time crop row detection method based on deep fusion image segmentation of spatial and contextual information, according to instructions stored in the memory.

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

[0031] Beneficial Effects: This application achieves spatial and contextual information extraction through a lightweight dual-branch structure. The low-resolution branch contains a context extractor, and multi-scale information is fused in parallel to achieve efficient extraction of long-range semantics. Utilizing the distinct channel information of crop growth cycle images, an information fusion module is constructed to perform deep, ordered fusion of the dual branches, removing useless information interference. This enables crop row detection for different crops, different growth stages, and different environments, achieving both accuracy and real-time performance. Attached Figure Description

[0032] Figure 1 A flowchart illustrating the steps of a crop row detection method;

[0033] Figure 2 This presents the overall framework of a crop row detection system based on deep fusion of spatial and contextual information for image segmentation.

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

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

[0036] Figure 5 This refers to a specific input image and its associated result image. Detailed Implementation

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

[0038] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following drawings denote similar items; therefore, once an item is defined in one drawing, it need not be further discussed in subsequent drawings.

[0039] Unless otherwise specified, the technical terms in this specification have the same meaning as those generally understood by those skilled in the art; however, in case of any conflict, the definitions in this specification shall prevail.

[0040] Therefore, the first technical solution of this application discloses a real-time crop row detection method based on deep fusion of spatial and contextual information for image segmentation, such as... Figure 1 As shown, it includes:

[0041] S1. Acquire the target image;

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

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

[0044] In this embodiment, the purpose of acquiring target images is to construct a crop row dataset to provide sufficient samples for later model training. In some preferred embodiments, the image acquisition method generally uses a DJI Osmo Mobile 6 gimbal, setting different pitch angles and acquisition heights to simulate the data acquisition states of different types of agricultural robots. This enhances the generalization performance of the algorithm trained on this dataset, enabling the algorithm to be directly deployed on various agricultural robots for row perception tasks. Furthermore, images are captured using a regular mobile phone at a resolution of 1920×1080 and a frame rate of 30fps, consistent with the image output format of commonly available inexpensive RGB cameras. The farmland row perception dataset covers various scene types, including different shooting angles, different crop types, different lighting conditions, different growth states, and other special cases. For shooting angles, data is acquired at different pitch and yaw angles. As the pitch angle decreases, the number of rows involved also decreases, covering the row detection needs of different agricultural robots. For crop types, we collected data on various common crops, including wheat, corn, eggplant, peppers, potatoes, soybeans, and cabbage. Regarding lighting conditions, we divided the data into daytime and nighttime. Different times of daytime have varying light intensities, while nighttime lighting was simulated using spotlights. For growth status, we collected data on the entire life cycle of wheat and corn. For other crops, we collected data on typical growth stages. Other factors included the presence of strong winds, ground moisture, crop row breaks, weeds, and other human-caused visual interference. The collected images were then labeled with crop rows using multi-line segments, assigning different line widths based on the crop's growth status, and converted into binary mask images.

[0045] In some specific implementations, the acquisition in S2, which involves deep fusion of spatial and contextual information of the target image, specifically includes:

[0046] (1) The target image is divided into spatial information branch and contextual information branch by a convolutional neural network;

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

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

[0049] In this embodiment, step (1) is: the target image is processed by a first convolutional layer with a 2x downsampling, and then processed by a second convolutional layer to further extract target image features, and the network is further differentiated into spatial information branches and context information branches.

[0050] In this embodiment, the first convolutional layer consists of 3x3 convolutions to achieve a 2x downsampling. The second convolutional layer transforms the feature size to 1 / 4 of the original image size, and further divides the network into spatial information branches and contextual information branches. Figure 2 As shown, the target image is divided into a spatial information branch and a contextual information branch after passing through a convolutional layer. To maintain high resolution and preserve spatial information, the convolution stride is set to 1, ensuring that the output size of the spatial information remains unchanged, thus maintaining high resolution. For the contextual information branch, the convolution stride is set to 2, making the image features smaller and smaller, thereby further extracting contextual information.

[0051] Furthermore, such as Figure 2 As shown, step (2) is as follows: the context information is processed by the third convolutional layer to achieve further information extraction. In addition, the spatial information is downsampled so that the spatial information features are the same size as the context information features. At this time, the spatial information plays a guiding role and is fused with the context information. The information is further extracted by the fourth convolutional layer to obtain the first context fusion information. The first context fusion information is upsampled so that the context information features are the same size as the spatial information features. At this time, the context information plays a guiding role and is fused with the spatial information. The information is further extracted by the fifth convolutional layer to obtain the first spatial fusion information. The above steps are repeated. After downsampling the first spatial fusion information, it plays a guiding role and is fused with the context information. The information is extracted by the sixth convolutional layer to obtain the second context fusion information. After upsampling the second context fusion information, it plays a guiding role and is fused with the spatial information. The information is extracted by the seventh convolutional layer to obtain the second spatial fusion information.

[0052] In this embodiment, spatial information is first fused with contextual information as a guide because, in the initial stage of differentiation, the feature resolution of contextual information is not low enough, resulting in insufficient extraction of high-level semantic information. Spatial information, however, maintains a higher resolution and retains more complete spatial information. Therefore, spatial information is unidirectionally input into contextual information to better capture contextual information. After the first fusion of spatial and contextual information, the first contextual fusion information is obtained through a convolutional layer. Its feature size is then reduced to 1 / 16 of the image size. At this point, the contextual information has acquired certain semantic information, which can provide some semantic information for the spatial information branch. Then, the contextual information is unidirectionally input into the spatial branch to obtain the first spatial fusion information, further enhancing spatial information capture. The above steps are then repeated to obtain the second contextual fusion information and the second spatial fusion information, ensuring that the spatial and contextual information of the image is fully extracted.

[0053] Furthermore, the fusion involves establishing different weight coefficients between channels by using a channel attention mechanism for spatial information and contextual information, then generating weights through a gating unit using a sigmoid function, and finally weighting the two parts of information to obtain the output.

[0054] The reason for assigning weights to different channels during fusion is understandable, as crops exhibit consistent color characteristics at different growth stages; for example, different types of crops are nearly green in the early stages of growth. Therefore, in the fusion stage, the dependencies between the two information features are first established using a channel attention mechanism. Figure 3 As shown, each channel is assigned a different weight coefficient, effectively strengthening important features and suppressing unimportant features. Furthermore, a gating unit fusion mechanism is designed, generating weights from the two inputs using a sigmoid function, and then weighting the two parts to obtain the output. The resulting context output incorporates spatial information, enabling better capture of contextual information in the downstream stages. It should be noted that the channel attention mechanism in the fusion steps of the entire process is consistent with that of 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. For example... Figure 4As shown, the second context fusion information is processed by the context information aggregation module to obtain aggregated context information. Specifically, the second context fusion information is used as input with a feature image at 1 / 32 image resolution. Pooling operations are performed in parallel: ① pooling kernel = 5, stride = 2; ② pooling kernel = 9, stride = 4; ③ pooling kernel = input feature resolution. This generates features at 1 / 64, 1 / 128, and 1 / 256 image resolutions. These features are then further processed by 1x1 convolution operations. The processed features are then added in parallel with the features obtained after the initial 1x1 convolution operation. The four features after parallel addition are then convolved with 3x3 convolution operations to achieve rapid aggregation of context information over a long distance, resulting in an efficient global representation.

[0056] In a further embodiment, the context aggregation information and the second spatial fusion information are merged to obtain the fused image features.

[0057] In some specific implementations, the image segmentation described in S3 is performed by inputting the fused image features into an image segmentation head. Specifically, during model training, an auxiliary segmentation head is added before the fifth convolutional layer to obtain an auxiliary cross-entropy loss, thereby optimizing the deep convolutional neural network. This auxiliary segmentation head L... 辅助 The generated auxiliary cross-loss is assigned a weight α, and added to the cross-entropy loss obtained from 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 during the testing phase, but only during the training phase. The model parameters are updated using the loss function to obtain the final segmentation model.

[0058] In some specific implementations, the line fitting in S3 is as follows: the segmented image is transformed into a bird's-eye view through the extrinsic lane line segmentation map, the number of crop rows is obtained by vertically accumulating and projecting the bird's-eye view, and then the points of the lane lines are searched using the sliding window method with the peak point as the starting point. Furthermore, the crop row lines are obtained by fitting the non-zero pixels in each sliding window using the least squares method.

[0059] It should be noted that the second to seventh convolutional layers in this application are all composed of blueprint-separated convolutional units. The number of blueprint-separated convolutional units, the number of output channels, and the stride of each pitch neural network can be determined according to requirements.

[0060] The second embodiment of this application discloses a crop row detection system based on deep fusion of spatial and contextual information for image segmentation. This system is used to implement the aforementioned crop row detection method based on deep fusion of spatial and contextual information for image segmentation, and includes:

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

[0062] Convolutional Neural Network Module: Includes convolutional neural network models one through seven, used to sample the target image or differentiated spatial and contextual information images; and

[0063] The fusion module includes a spatial information fusion module, a contextual information fusion module, a channel attention mechanism module for weight allocation during fusion, and a gating unit module, which are used to achieve effective fusion of spatial and contextual 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 and prediction of the fused image features.

[0066] The present invention will be specifically described below through examples.

[0067] Figure 5 The first row shows corn images from a region in Beijing. The images vary in shooting angle, height, light intensity, and crop growth stage, highlighting the richness and challenge of the dataset. The second row shows the corresponding labeled images. Crop rows were marked with multi-line segments using a labeling tool, with different line widths assigned based on the crop's growth stage, and then converted into binary mask images. The third row shows the segmentation results obtained using the model trained with the described method, compared to the labeled results in the second row, demonstrating better segmentation performance. The fourth row is a bird's-eye view of the segmented image after extrinsic perspective transformation. The fifth row is the result after line fitting and inverse perspective transformation, showing good alignment between the fitted image and the center of the crop rows.

[0068] The table below shows the quantitative test results on the self-collected dataset. It can be seen that this algorithm maintains high accuracy compared with Unet and Fast-SCNN while having the fastest computing speed, realizing real-time detection of crop rows.

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

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

[0071] Memory; and

[0072] The processor is configured to execute the above-described real-time crop row detection method based on deep fusion image segmentation of spatial and contextual information, according to instructions stored in the memory.

[0073] Finally, this application also discloses a computer-readable storage medium storing computer instructions that, when executed by a processor, can realize the above-described real-time crop row detection method based on deep fusion image segmentation of spatial and contextual information.

[0074] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A real-time crop row detection method based on deep fusion of spatial and contextual information for image segmentation, comprising: S1. Acquire the target image; S2. Based on a convolutional neural network, perform deep fusion of spatial and contextual information on the target image; S3. The fused image features are classified and predicted at the pixel level using a segmentation head, and the segmented results are fitted with lines to obtain crop rows; Specifically, S2 describes the deep fusion of spatial and contextual information of the target image: (1) The target image is divided into spatial information branches and contextual information branches by a convolutional neural network; (2) Deeply fuse the spatial and contextual information of the two branches; (3) Merge the fused spatial information and contextual information to obtain the fused image features; Step (2) is as follows: The context information is processed by the third convolutional layer to achieve further information extraction. In addition, the spatial information is downsampled so that the spatial information features are the same size as the context information features. At this time, the spatial information plays a guiding role and is fused with the context information. The information is further extracted by the fourth convolutional layer to obtain the first context fusion information. The first context fusion information is upsampled so that the first context information features are the same size as the spatial information features. At this time, the first context information plays a guiding role and is fused with the spatial information. The information is further extracted by the fifth convolutional layer to obtain the first spatial fusion information. The above steps are repeated. After downsampling the first spatial fusion information, it plays a guiding role and is fused with the first context information. The information is extracted by the sixth convolutional layer to obtain the second context fusion information. After upsampling the second context fusion information, it plays a guiding role and is fused with the first spatial information. The information is extracted by the seventh convolutional layer to obtain the second spatial fusion information. The obtained second context fusion information is aggregated to obtain context aggregation information, thereby obtaining an efficient global representation of the context information. The context aggregation information and the second spatial fusion information are merged to obtain the fused image features.

2. The method according to claim 1, characterized in that, The step (1) is as follows: the target image is processed by a first convolutional layer with a 2x downsampling, and then processed by a second convolutional layer for further feature extraction, and the network is further divided into spatial information branches and contextual information branches.

3. The method according to claim 1, characterized in that, The fusion process involves establishing different weight coefficients between channels using a channel attention mechanism for spatial and contextual information, then generating weights through a gating unit using a sigmoid function, and finally weighting the two parts of information to obtain the output.

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

5. The method according to claim 1, characterized in that, The line fitting described in S3 is as follows: the segmented image is transformed into a bird's-eye view through extrinsic parameters, the number of crop rows is obtained by vertically accumulating and projecting the bird's-eye view, and then the points of the lane lines are searched using the sliding window method with the peak point as the starting point. Furthermore, the crop row lines are obtained by fitting the non-zero pixels in each sliding window using the least squares method.

6. A crop row detection system based on deep fusion of spatial and contextual information for image segmentation, used to implement the method as described in any one of claims 1-5, comprising: Image acquisition module: used to acquire target images; Convolutional Neural Network Module: Includes the first to seventh convolutional neural network models, used to extract features from the target image or the differentiated spatial and contextual information image; and Fusion Module: Includes a spatial information fusion module and a contextual information fusion module, as well as a channel attention mechanism module and a gating unit module for weight allocation during fusion, used to achieve effective fusion of spatial and contextual 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 and prediction of the fused image features.

7. A computer device, characterized in that, include: Memory; And a processor configured to execute, based on instructions stored in the memory, the real-time crop row detection method based on spatial and contextual information deep fusion image segmentation as described in any one of claims 1-5.

8. A computer-readable storage medium storing computer instructions that, when executed by a processor, enable the real-time crop row detection method based on deep fusion image segmentation using spatial and contextual information as described in any one of claims 1-5.

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