A method, device, medium and product for monitoring corn seeding quality

Abstract

The application discloses a corn seeding quality monitoring method and device, medium and product, relates to the field of agricultural intelligence and computer vision technology, and the method comprises the following steps: acquiring a corn seeding video frame sequence containing motion blur; performing deblurring processing on the corn seeding video frame sequence through a pre-trained super-resolution deblurring network to obtain a clear image sequence; identifying the clear image sequence through a pre-trained target detection network to obtain a corn seed target sequence; accumulating the change relationship between the cumulative number of the corn seed target sequence passing through a pre-set detection line and time, and comparing the change relationship with a reference relationship to determine corn seed over-seeding events and missing seeding events. The application realizes high-precision identification and positioning of corn seeds in degraded frames, improves detection accuracy, and effectively supports reliable determination of over-seeding and missing seeding events.

Description

Technical Field

[0001] This application relates to the fields of agricultural intelligence and computer vision technology, and in particular to a method, equipment, medium and product for monitoring the quality of corn planting. Background Technology

[0002] During corn planting, even seed distribution is crucial for ensuring crop yield and healthy plant growth. When seeders operate in the field, variations in terrain, equipment vibration, and operating speed often lead to double-seeding or missed seeding, resulting in uneven spatial distribution of plants. This intensifies competition for water and fertilizer among individual plants, ultimately causing yield reduction. Therefore, real-time monitoring of planting quality has significant agricultural application value.

[0003] Existing vision-based methods for monitoring sowing quality typically assume that the acquired images are clear, non-degrading still images, requiring stable detection equipment and good lighting conditions. However, during field sowing, the seeder is in high-speed motion, and the severe vibrations caused by uneven road surfaces result in motion blur and ghosting in the acquired video frames. Existing methods heavily rely on clear, still images, making them ill-suited to the dynamic blurring and degradation environments prevalent in actual field operations. This leads to excessively high requirements for imaging conditions and hardware stability, resulting in significant performance degradation in practical deployments.

[0004] Corn seeds occupy only a few pixels in an image, making them typical small targets. Motion blur further erases high-frequency details such as edges and textures of the seed targets, exacerbating the already difficult problem of small target detection. Motion blur causes the loss of high-frequency details in corn seeds, severely suppressing the ability to detect small targets; seeds in degraded frames are almost unrecognizable by existing detection models. Some studies have attempted to introduce image deblurring or super-resolution reconstruction as preprocessing steps, but existing restoration methods are mostly designed for human visual perception quality and lack targeted enhancement of high-frequency discriminative features in downstream detection tasks. Furthermore, the restoration and detection modules are independent and lack effective collaborative optimization, resulting in residual blur artifacts in the restored image and limited improvement in detection accuracy. Existing technologies lack mechanisms for explicitly enhancing edge and texture information in the frequency domain, and also lack a collaborative detection architecture combining attention and multi-scale fusion, making it difficult to achieve high-precision identification and localization of corn seeds under degraded conditions. Summary of the Invention

[0005] The purpose of this application is to provide a method, device, medium, and product for monitoring the quality of corn sowing, which can directly access and process degraded video streams acquired in real time under complex operating environments, reducing the requirements for imaging conditions and hardware stability; and achieve high-precision identification and positioning of corn seeds in degraded frames, so that the detection accuracy of degraded images that were originally almost undetectable is improved after processing, effectively supporting the reliable determination of replanting and missed sowing events.

[0006] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a method for monitoring maize planting quality, including: Obtain a sequence of video frames containing motion blur from corn planting; The corn sowing video frame sequence is deblurred using a pre-trained super-resolution deblurring network to obtain a clear image sequence. The super-resolution deblurring network includes a wavelet enhancement residual module, which performs wavelet domain decomposition and enhancement processing on the input features. The clear image sequence is identified by a pre-trained target detection network to obtain a corn seed target sequence. The target detection network includes a composite attention mechanism module and a multi-scale fusion module. The composite attention mechanism module combines improved coordinate attention, efficient channel attention, and parameter-free spatial attention. The multi-scale fusion module performs multi-scale feature fusion through multiple parallel convolutional branches with different dilation rates. The cumulative number of corn seed target sequences passing through a preset detection line is accumulated over time, and the relationship between the cumulative number of these sequences and a baseline is compared to determine whether a corn seed replanting event or a missed replanting event occurs.

[0007] Optionally, the training process of the pre-trained super-resolution deblurring network includes: A randomized multi-stage degradation dataset is constructed, the construction process of which includes: simulating motion blur by averaging the pixels of consecutive frames from high-speed sampled clear video; and randomly applying multiple degradation stages to the generated clear-blurred image pairs, each degradation stage randomly including one or more operations among optical blur, optical chromatic aberration, sensor noise, quantization, compression artifacts, and downsampling. Using the sharp image in the sharp-blurred image pair as the supervision signal, the super-resolution deblurring network is trained on the random multi-stage degradation dataset until convergence.

[0008] Optionally, the training process of the pre-trained object detection network includes: This is achieved through two-stage training, which includes: In the first training phase, small local image patches centered on the target box are cropped from the original labeled image for training to obtain the optimal weights. In the second training phase, the object detection network is initialized with the optimal weights and training continues on the original labeled images.

[0009] Optionally, the specific process of performing wavelet domain decomposition and enhancement processing on the input features using the wavelet-enhanced residual module includes: Perform a two-dimensional discrete wavelet transform of the Haar wavelet basis on the input features to obtain a low-frequency sub-band and three high-frequency sub-bands in three directions, namely, a horizontal high-frequency sub-band, a vertical high-frequency sub-band, and a diagonal high-frequency sub-band. The low-frequency subband is subjected to feature transformation by 1×1 convolution and residual connection to obtain the enhanced low-frequency subband. The high-frequency subbands in the three directions are spliced ​​together in the channel dimension, and then adaptively weighted and enhanced by depthwise separable convolution and channel attention mechanism in sequence. Finally, residual connection is performed with the high-frequency subbands before splicing to obtain the enhanced high-frequency subbands. The enhanced low-frequency subband and the enhanced high-frequency subband are reconstructed into output features through inverse discrete wavelet transform.

[0010] Optionally, the pre-trained super-resolution deblurring network further includes a three-stage progressive upsampling module, which receives low-resolution features output by the encoder in the super-resolution deblurring network as input, and includes: In the first upsampling stage, the low-resolution features are upsampled by a factor of two, and the low-frequency signal is enhanced by the wavelet enhancement residual module to restore the overall structure of the image. In the second upsampling stage, the features output in the first stage are upsampled by a factor of two again, and the low-frequency and high-frequency signals are simultaneously enhanced by the wavelet enhancement residual module to restore texture details. In the third upsampling stage, the high-frequency signal of the features output from the second stage is enhanced by the wavelet enhancement residual module to sharpen the image edges.

[0011] Optionally, the composite attention mechanism module includes: The improved coordinate attention introduces a minimum channel number constraint during the channel compression stage and employs group normalization and the SiLU activation function. The efficient channel attention uses one-dimensional convolution with a fixed kernel size; The parameterless spatial attention introduces a numerically stable term into the variance term of the neuron energy calculation to smooth the attention weights. The composite attention mechanism module uses the improved coordinate attention as the backbone and fuses the output of the high-efficiency channel attention and the parameterless spatial attention in series with the output of the improved coordinate attention.

[0012] Optionally, the multi-scale fusion module includes three parallel convolutional branches: a small receptive field branch, a medium receptive field branch, and a large receptive field branch, wherein the three parallel convolutional branches are allocated with asymmetric channels.

[0013] Secondly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the corn planting quality monitoring method described in any one of the above.

[0014] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the corn planting quality monitoring method described above.

[0015] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the corn planting quality monitoring method described above.

[0016] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a method, device, medium, and product for monitoring corn planting quality. By acquiring a sequence of corn planting video frames containing motion blur, it solves the problem of dynamic blur degradation caused by the reliance on static, clear images in existing methods, which are difficult to adapt to the high-speed movement and shaking of field planters. It enables direct access and processing of degraded video streams acquired in real time under complex operating environments, reducing the requirements for imaging conditions and hardware stability. The video frame sequence is deblurred using a pre-trained super-resolution deblurring network, and the cleared image sequence is identified using a pre-trained target detection network. This solves the coupling technical problem of loss of high-frequency details of corn seeds and severe suppression of small target detection ability caused by motion blur, achieving high-precision identification and positioning of corn seeds in degraded frames. Specifically, a wavelet enhancement residual module is used to explicitly enhance edge and texture information in the frequency domain, and a composite attention mechanism and multi-scale fusion module are used to improve the feature capture ability of small targets. The synergy of these two methods improves the detection accuracy of degraded images that were originally almost undetectable after processing, effectively supporting the reliable determination of replanting and missed planting events. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is an application environment diagram of a corn planting quality monitoring method according to an embodiment of this application; Figure 2A schematic flowchart illustrating a method for monitoring corn planting quality according to an embodiment of this application; Figure 3 This is a schematic diagram of the network structure of the target detection network in one embodiment of this application; Figure 4 This is a schematic diagram of the workflow of the wavelet enhancement residual module in one embodiment of this application; Figure 5 This is a schematic diagram of the workflow of a three-stage progressive upsampling module in one embodiment of this application; Figure 6 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] The corn planting quality monitoring method provided in this application embodiment can be applied to, for example... Figure 1In the application environment shown, the image acquisition terminal 102 (such as an industrial camera) installed on the sowing equipment communicates with the server 104 via a network. The data storage system can store the video data that the server 104 needs to process, trained network model parameters, and monitoring results. The data storage system can be set up separately, integrated into the server 104, or placed in the cloud or on another server. The image acquisition terminal 102 can send the real-time acquired sowing operation video stream to the server 104 via the network. After receiving the video stream, the server 104, for each frame or consecutive frame sequence in the video stream, performs deblurring processing on the video frames containing motion blur based on a pre-trained super-resolution deblurring network to obtain a clear image sequence; then, it uses a pre-trained target detection network to identify and locate the corn seed targets in the clear image sequence to obtain a corn seed target sequence; the server 104 automatically accumulates the number of targets passing through preset detection lines according to the target sequence and generates a relationship of change over time, compares it with a benchmark relationship, and determines replay events and missed sowing events. Server 104 can provide feedback on the judgment results (such as timestamps, locations, statistical charts, etc. of replay / missed sowing) to image acquisition terminal 102 or the cockpit display terminal of the sowing equipment. Furthermore, in some embodiments, the corn sowing quality monitoring method can also be executed independently by the edge computing module built into image acquisition terminal 102. For example, the onboard industrial control computer on the sower can directly process and judge the acquired video stream in real time, or server 104 can retrieve pre-recorded sowing operation videos from the data storage system for offline analysis.

[0022] The image acquisition terminal 102 can be, but is not limited to, various industrial cameras, high-speed cameras, or dedicated agricultural vision sensors. The vehicle-mounted industrial control computer on the seeding equipment can be an embedded computer, a programmable logic controller (PLC), or an edge computing box. The server 104 can be implemented using a standalone server, a server cluster consisting of multiple servers, or a cloud server.

[0023] In one exemplary embodiment, such as Figure 2 As shown, a method for monitoring corn planting quality is provided. This method is executed by computer equipment, specifically by a single computer device such as an image acquisition terminal 102, a vehicle-mounted industrial control computer, or a server 104, or by both the image acquisition terminal 102 and the server 104. In this embodiment, the method is applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps 201 to 204. Wherein: Step 201: Obtain a sequence of corn planting video frames containing motion blur; Step 202: The corn sowing video frame sequence is deblurred using a pre-trained super-resolution deblurring network to obtain a clear image sequence. The super-resolution deblurring network includes a wavelet enhancement residual module, which performs wavelet domain decomposition and enhancement processing on the input features. Step 203: The clear image sequence is identified by a pre-trained target detection network to obtain the corn seed target sequence. The target detection network includes a composite attention mechanism module and a multi-scale fusion module. The composite attention mechanism module combines improved coordinate attention, efficient channel attention and parameter-free spatial attention. The multi-scale fusion module performs multi-scale feature fusion through multiple parallel convolution branches with different dilation rates. Step 204: Accumulate the cumulative number of corn seed target sequences over time using a preset detection line, compare the changes with a baseline, and determine whether the corn seed replanting event or the missed replanting event is a replanting event.

[0024] Implementing steps 201 to 204 above solves the problem of dynamic blurring degradation caused by the reliance on static, clear images in existing methods, which is difficult to adapt to the high-speed movement and shaking of field seeders. It enables direct access and processing of degraded video streams acquired in real time under complex operating environments, reducing the requirements for imaging conditions and hardware stability. It also solves the coupling technical problem of loss of high-frequency details in corn seeds and severe suppression of small target detection ability caused by motion blur, achieving high-precision identification and positioning of corn seeds in degraded frames. In particular, the wavelet enhancement residual module is used to explicitly enhance edge and texture information in the frequency domain, and the composite attention mechanism and multi-scale fusion module are used to improve the feature capture ability of small targets. The synergy of the two makes the detection accuracy of degraded images that were originally almost undetectable improved after processing, effectively supporting the reliable determination of replay and missed replay events.

[0025] As an optional implementation, the training process of the pre-trained super-resolution deblurring network includes: A randomized multi-stage degradation dataset is constructed. The construction process includes: simulating motion blur by averaging the pixels of consecutive frames from high-speed sampled sharp videos; and randomly applying multiple degradation stages to the generated sharp-blurred image pairs. Each degradation stage randomly includes one or more operations among optical blur, optical chromatic aberration, sensor noise, quantization, compression artifacts, and downsampling. Using the sharp image in a sharp-blurred image pair as the supervision signal, the super-resolution deblurring network is trained on a random multi-stage degradation dataset until convergence.

[0026] Specifically, firstly, the original frame rate of the live-action sowing video is increased to 1920fps using the RIFE frame interpolation algorithm to capture the motion trajectory of corn seeds during the sowing process with high temporal resolution. One hundred segments are extracted from this high frame rate video, each containing 100 consecutive frames as a sharp reference. Then, for each video segment, the motion blur time window length is calculated based on the target video frame rate and a duty cycle of 0.8. Within this time window, consecutive frames are directly averaged using RGB pixel space to simulate the real exposure process, generating corresponding motion blur images, thus forming sharp-blurred image pairs.

[0027] To simulate the complex and variable degradation environment in the field, a random degradation process consisting of 1 to 3 degradation stages is applied to the generated clear-blurred image pairs. The degradation parameters for each stage are randomly generated based on a global configuration. Specifically, in the first stage, optical blur (such as Gaussian blur) is applied, and optical chromatic aberration shift is randomly added according to a preset probability. Simultaneously, sensor noise is superimposed and quantized to simulate the real optical imaging and sensor acquisition process. In the second and third stages, the image is further blurred using randomly generated anisotropic Gaussian convolution kernels, and sensor noise and quantization operations are added again. If compression is selected to be applied in any degradation stage (with a fixed probability of 40%), video compression and decoding simulation is first applied to the frame sequence to generate compression artifacts, and then bicubic interpolation is uniformly used for downsampling. Clear images are directly downsampled without compression artifact processing. Through the above process, each video segment can generate diverse blind degradation samples, effectively reducing the model's dependence on a single degradation mode, enabling the trained network to adapt to multiple coupled degradations such as motion blur, noise, and compression distortion that occur in the actual sowing process.

[0028] Using the sharp image from the sharp-blurred image pair as the supervision signal, the corresponding blurred image is input into the super-resolution deblurring network (i.e., WER-FMA-Net) to be trained. The network outputs the restored image, and the loss function between the output image and the sharp image is calculated (L1 loss, perceptual loss, or combined loss can be selected). The Adam optimizer is used, with an initial learning rate of 1e-4 and a batch size of 8, and iterative training is performed until the loss function converges. During training, the network extracts multi-scale features of the blurred image through the encoder, enhances high-frequency details in the frequency domain using the wavelet-enhanced residual module (WER), and gradually restores the image resolution through a three-stage progressive upsampling module. The final pre-trained model can effectively remove motion blur, enhance edge and texture information, and provide high-quality input images for downstream object detection tasks.

[0029] As an optional implementation, the training process of the pre-trained object detection network includes: This was acquired through a two-stage training program, which included: In the first training phase, small local image patches centered on the target box are cropped from the original labeled image for training to obtain the optimal weights. In the second training phase, the object detection network is initialized with optimal weights and training continues on the original labeled images.

[0030] Specifically, in the first training phase, a 300×300 pixel cropping region is generated outward from the center of each labeled target bounding box. If the cropping region exceeds the image boundary, it is filled with blank pixels. 1000 such target bounding box centers are randomly selected from the training data to generate corresponding local small image patches, constituting a simplified dataset for the first training phase. This simplified dataset is then input into the object detection network (the baseline network is YOLOv8n, which embeds a DSM-HA module, and its network structure is as follows). Figure 3 (As shown). The training hyperparameters are set as follows: input resolution 300×300, initial learning rate lr0=0.001, final learning rate lrf=0.0001, weight decay=0.0001, momentum=0.937. A stochastic gradient descent optimizer is used, and training is performed for 50 epochs. The goal of this training stage is to allow the network to focus on learning fine-grained features of the corn seed itself, including discriminative information such as seed shape, edges, and texture, avoiding interference from complex backgrounds. After training converges, the optimal weights for this stage are saved.

[0031] In the second training phase, the optimal weights obtained in the first training phase are used as the initialization parameters for the object detection network. Complete, original resolution labeled images (1024×1024 resolution) are input into the network, and training continues on the complete dataset. The training hyperparameters are set as follows: input resolution 1024×1024, initial learning rate lr0=0.01, final learning rate lrf=0.01, weight decay=0.0005, and momentum=0.937. A stochastic gradient descent optimizer is used, and training is performed for 100 epochs. The goal of this training phase is to enable the network to further adapt to complex field background environments while maintaining its sensitivity to local object details, thus enhancing its ability to model global contextual information. After training, the object detection network obtained in the second training phase is used as the final pre-trained object detection network.

[0032] It should be noted that the two training stages mentioned above work synergistically with the DSM multi-scale fusion module in the object detection network. The DSM module covers different receptive fields through parallel convolutional branches with multiple dilation rates: the small receptive field branch focuses on local details of the target, the medium receptive field branch captures medium semantic regions, and the large receptive field branch integrates global contextual information. In the first stage of training with cropped small images, the network extracts fine-grained features of the target using the small receptive field; in the second stage of training with the complete image, the small receptive field branch still maintains accurate capture of local target features, while the medium and large receptive field branches can fuse the spatial relationship between the target and the complex background. This design enables the network to stably extract target features under different contextual conditions, taking into account both local details and global semantics, effectively improving the robustness and detection accuracy of the model in complex field environments. In both training stages, the network's default data augmentation methods are used to perform online enhancement of the input image, including but not limited to random flipping, random scaling, and color dithering, to further improve the model's generalization ability and robustness.

[0033] As an optional implementation, the specific process of performing wavelet domain decomposition and enhancement processing on the input features using the wavelet-enhanced residual module includes: Perform a two-dimensional discrete wavelet transform of the Haar wavelet basis on the input features to obtain a low-frequency subband and three high-frequency subbands in three directions, namely the horizontal high-frequency subband, the vertical high-frequency subband, and the diagonal high-frequency subband. The low-frequency subband is enhanced by performing feature transformation through 1×1 convolution and residual connection. The high-frequency subbands in the three directions are spliced ​​together along the channel dimension, and then adaptively weighted and enhanced by depthwise separable convolution and channel attention mechanism. Finally, residual connections are made with the high-frequency subbands before splicing to obtain the enhanced high-frequency subbands. The enhanced low-frequency subband and enhanced high-frequency subband are reconstructed into output features through inverse discrete wavelet transform.

[0034] Specifically, let the input feature mapping be... Where B represents the batch size, C represents the number of channels, and H and W represent the height and width of the feature map, respectively. A single-layer two-dimensional Haar wavelet decomposition is performed independently on each channel of the input feature X. The two-dimensional Haar wavelet can be viewed as performing one-dimensional wavelet filtering and downsampling in the row and column directions, respectively. Its low-pass filter g and high-pass filter h are defined as follows: ; .

[0035] The input feature X is first filtered and downsampled in the horizontal direction, and then the same operation is performed in the vertical direction. This yields one low-frequency sub-band and three high-frequency sub-bands, corresponding to the approximate information LL and detailed information in the horizontal (LH), vertical (HL), and diagonal (HH) directions, respectively. Specifically, wavelet decomposition can be written as... ; in, It is the low-frequency subband LL.

[0036] The high-frequency subbands LH, HL, and HH in the three directions are represented as follows: .

[0037] From a pixel-level perspective, the calculation of DWT in the single-channel case can be formally expressed as: ; ; ; ; in This represents the position index after downsampling. This decomposition process, while preserving energy reversibility, explicitly splits the input features into global structural information and orientation-sensitive high-frequency detail information.

[0038] In both mode, the low-frequency components are primarily responsible for characterizing the overall contour and brightness distribution of the image. To make appropriate adjustments without disrupting the global structure, we employ lightweight residual mapping for the low-frequency sub-band: ; Among them, 1×1 convolution is used for linear recombination of the channel dimension, residual connection ensures stable transmission of low-frequency information, and PReLU activation introduces the necessary nonlinear expressive power.

[0039] The high-frequency subbands primarily carry edge, texture, and detail structure information. In both mode, the high-frequency components in the three directions are first merged along the channel dimension for unified modeling. ; Subsequently, local texture responses are extracted using depthwise separable convolution: ; And a channel attention mechanism is introduced to generate adaptive weights: ; ; in,( () represents the Sigmoid function. Ultimately, high-frequency enhancement is achieved in the form of residuals: ; The enhanced high-frequency features are then split along the directional dimension to restore the original sub-band structure: .

[0040] After completing the independent modeling of the low-frequency and high-frequency subbands, the frequency domain features are reconstructed back to the spatial domain through inverse discrete wavelet transform: .

[0041] For Haar wavelets, IDWT can be considered the strict inverse of DWT, essentially involving upsampling of the low-frequency and high-frequency subbands and weighted summation using corresponding reconstruction filters. In the single-channel case, the reconstruction formula can be written as: ; in,( This process ensures the natural integration of low-frequency global structure and enhanced high-frequency details without information loss.

[0042] The final output features are: .

[0043] This output feature contains both stable global semantic information and adaptively enhanced local detail structure, and can be directly used for subsequent image restoration, enhancement, or reconstruction tasks. The specific workflow is as follows: Figure 4 As shown.

[0044] As an optional implementation, the pre-trained super-resolution deblurring network further includes a three-stage progressive upsampling module, which receives low-resolution features output by the encoder in the super-resolution deblurring network as input, and includes: In the first upsampling stage, the low-resolution features are upsampled by a factor of two, and the low-frequency signal is enhanced by a wavelet-enhanced residual module to restore the overall structure of the image. In the second upsampling stage, the features output from the first stage are upsampled by a factor of two again, and the low-frequency and high-frequency signals are simultaneously enhanced by the wavelet enhancement residual module to restore texture details. In the third upsampling stage, the high-frequency signal of the features output from the second stage is enhanced by a wavelet-enhanced residual module to sharpen the image edges.

[0045] Specifically, let the low-resolution features output by the encoder be... (That is, it is downsampled by 4 times relative to the original input image resolution), and this low-resolution feature is used as the input of the three-stage progressive upsampling module.

[0046] In the first upsampling phase, the low-resolution features Perform double upsampling to obtain a resolution of The intermediate features are obtained. This upsampling process can use a dynamic upsampling module (such as the duf module in FMA-Net) to fuse the features after optical flow warping with the residual features. The upsampled features are input into the wavelet-enhanced residual module and configured in low-frequency mode: only the low-frequency subband is enhanced, while the high-frequency subband remains unchanged. In this mode, the wavelet-enhanced residual module extracts the overall contour and brightness distribution information of the image through discrete wavelet transform, enhances the low-frequency components through 1×1 convolution and residual concatenation, and then reconstructs them through inverse wavelet transform. The output features at this stage are... The resolution is Furthermore, it is combined with the double residual features in the super-resolution deblurring network. Element-by-element addition and fusion are performed to enhance the overall structural information of the image.

[0047] The features output in the first upsampling stage As input to the second upsampling stage, the feature is upsampled again by a factor of two to obtain a resolution of H×W. The upsampled feature is then input to the wavelet enhancement residual module and configured in Both mode: simultaneously enhancing both the low-frequency and high-frequency subbands. In this mode, the module enhances the stability of the low-frequency subband and adaptively weights the high-frequency subbands in three directions (horizontal, vertical, and diagonal) using channel attention and depth-separable convolution to recover details such as edges and textures. This stage outputs the feature. The resolution is H×W, and it is compared with the four-fold residual features in the super-resolution deblurring network. Element-by-element addition and fusion are performed to further improve the quality of texture detail restoration.

[0048] The features output from the second upsampling stage The input is fed into the wavelet enhancement residual module and configured in high-frequency mode: only the high-frequency subband is enhanced, while the low-frequency subband remains unchanged. In this mode, the module focuses on enhancing high-frequency details in the horizontal, vertical, and diagonal directions, further sharpening image edges and making the boundaries of the output image clearer and more continuous. The output features at this stage... The resolution is H×W. The features output from the third upsampling stage... Again with the aforementioned four times residual feature Element-wise addition and fusion are performed to obtain the final output feature. This output feature is then mapped to the RGB space through a convolutional layer (e.g., a 3×3 convolution) to generate the restored, clear image. The specific workflow is as follows: Figure 5 As shown.

[0049] As an optional implementation, the composite attention mechanism module includes: The improved coordinate attention introduces a minimum channel number constraint in the channel compression stage and employs group normalization and SiLU activation function; Efficient channel attention uses one-dimensional convolution with a fixed kernel size; Parametric spatial attention introduces a numerically stable term into the variance term used to calculate neuron energy in order to smooth the attention weights. The composite attention mechanism module uses improved coordinate attention as its backbone and fuses the output of efficient channel attention and parameterless spatial attention in series with the output of improved coordinate attention.

[0050] Specifically, the improved CoordAttention-E, based on standard CoordAttention, introduces a minimum channel number constraint and employs Group Normalization and the SiLU activation function. First, it performs global average pooling on the input feature X along both the horizontal and vertical directions to embed coordinate information: ; ; , By concatenating the feature maps from the two directions, we obtain... Dimensionality reduction is then performed using shared 1×1 convolutions. During the channel compression stage, a minimum channel number constraint is introduced: the number of channels after compression. Where r is the compression ratio (usually set to 16 or 32), ensuring that the number of channels after compression is not less than 4, avoiding excessive reduction of key information. The dimensionality-reduced features are normalized by group normalization (the number of groups can be set to 32), and then activated by the SiLU activation function: ; Will Separate into two independent tensors along the horizontal and vertical directions. and The channels are restored to their original number C through 1×1 convolutions, and then the horizontal attention weights are generated by the Sigmoid function. and vertical attention weights : , ; Finally, the output of the improved coordinate attention is: .

[0051] Efficient Channel Attention (ECA-F) uses one-dimensional convolutions with a fixed kernel size to model local dependencies across channels. First, global average pooling is performed on the input feature X to obtain the channel description vector. Then, a one-dimensional convolution (with a fixed kernel size of k=3) is used to process the vector to learn the dependencies between each channel and its neighboring channels: ; The Conv1D padding is set to ensure the output length is the same as the input. The convolution result is activated using a Sigmoid function to obtain the channel attention weights. ; The output of efficient channel attention is: .

[0052] Parametric Spatial Attention (SimAM-S) constructs spatial attention weights based on the neuron energy function and introduces a numerically stable term into the variance term. (Pick =0.1) to smooth attention weights.

[0053] For the input feature X, calculate the spatial location. The energy function of the neuron at position c. For position c in channel c... neurons at the location The difference between its response and the average response at other spatial locations within the same channel can be measured by an energy function. The minimum expression for the energy function is: ; in, Let be the mean of channel c. Let be the variance of channel c. To avoid statistical degradation of variance estimation at low resolution or in response sets, a numerical stabilization term is introduced. Smoothing constraints are applied to the degenerate variance.

[0054] Energy value per neuron The lower the value, the more significant the difference between the neuron and other locations in its channel, and the more important the corresponding location. Inverting the energy value and activating it via a sigmoid function yields the spatial attention weights: ; The output of parameterless spatial attention is: ; This process requires no additional learnable parameters and is computationally efficient.

[0055] The composite attention mechanism module uses improved coordinate attention as the main branch, and combines efficient channel attention and parameterless spatial attention as auxiliary branches.

[0056] Main path: Input feature X undergoes improved coordinate attention processing, output... .

[0057] Auxiliary path: Input feature X first undergoes efficient channel attention processing, output... Then As input, after parameterless spatial attention processing, the output is... .

[0058] Finally, the main output and the auxiliary path outputs are added together to obtain the final output of the composite attention mechanism: .

[0059] As an optional implementation, the multi-scale fusion module includes three parallel convolutional branches: a small receptive field branch, a medium receptive field branch, and a large receptive field branch, with asymmetric channels assigned to the three parallel convolutional branches.

[0060] Specifically, asymmetric channels: Small receptive field branch (d1 branch): Allocates more channels to capture local texture and detail features. Let the number of channels allocated to this branch be... ,in A value of 0.5 can be used (i.e., half of the channels are reserved).

[0061] The middle receptive field branch (d3 branch): allocates fewer channels to capture information from medium semantic regions. Let the number of channels allocated to this branch be... ,in 0.25 is acceptable.

[0062] Large receptive field branch (d5 branch): Allocates fewer channels to capture global context information. Let the number of channels allocated to this branch be... ,in A value of 0.25 can be taken, which is the same as the middle receptive field branch.

[0063] The above allocation satisfies .

[0064] The small receptive field branch uses a simple 3×3 convolution (dilation=1) to directly extract local detail features without introducing additional complex operations to maintain the integrity of the detailed information. The output is denoted as... The receptive field branch first undergoes depthwise convolution (3×3, dilation=3), followed by pointwise convolution (1×1). After convolution, it is sequentially processed by batch normalization and SiLU activation function to refine features and suppress redundant information. The output is denoted as... The large receptive field branch first undergoes depthwise convolution (3×3, dilation=5), followed by pointwise convolution (1×1), then batch normalization and SiLU activation. To further filter key channels and spatial regions, the composite attention mechanism module (HA) can be introduced after activation to perform attention-weighted processing on the output features. The output is denoted as... .

[0065] ; The concatenated features are then compressed back to the original number of channels C using a 1×1 convolution: .

[0066] The fused features The element-wise addition with the original input feature X yields the final output of the multi-scale fusion module: .

[0067] Residual connections ensure stable gradient propagation and enable the network to learn identity mappings, thus preventing degradation of deep networks.

[0068] This application also provides an application scenario in which the aforementioned corn planting quality monitoring method is applied. Specifically, the corn planting quality monitoring method provided in this embodiment can be applied to intelligent quality control scenarios in corn planting operations on large farms. After the planting operation begins, an industrial camera installed behind the seeder collects video streams of the planting furrows in real time. After deblurring and target detection processing, the video stream generates a seed passing count curve with a timestamp. The system compares the actual count curve with the theoretical curve calculated based on the seeder's travel speed and nominal plant spacing in real time. Once a reseeding or missed seeding event is detected, an audible and visual alarm is immediately issued on the display terminal in the cockpit, and the abnormal location is marked in the geographic information system. Simultaneously, all monitoring data is uploaded to the farm management cloud platform via a 4G / 5G network to form a planting operation quality report, providing data support for subsequent reseeding and agricultural machinery operation evaluation. This method belongs to the core monitoring link in the digital closed-loop management of planting operations. Specifically, through the collaboration of machine vision and deep learning models, it replaces traditional manual sampling and sensor monitoring methods, realizing real-time, accurate, and non-contact evaluation of planting quality.

[0069] It should be noted that, to achieve seeding detection, this application constructs a complete method for monitoring maize seeding quality. Based on the international standard ISO 7256-1:198431, the nominal inter-particle spacing of the seeder is determined. Used to determine seeding quality (benchmark relationship): when the intergranular spacing It was a rebroadcast. It was a missed broadcast at the time, and This is considered normal sowing. When the seeder is operating at a constant speed, The distance between seeds is fixed, so the interval between seed appearances can be calculated. In practical applications, it can be arbitrarily set according to operational requirements. This method has no restrictions on this as an input parameter.

[0070] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 6 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores video data to be processed, trained network model parameters, and monitoring results. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a method for monitoring corn planting quality.

[0071] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0072] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0073] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0074] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0075] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0076] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0077] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0078] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0079] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for monitoring the quality of corn planting, characterized in that, The method for monitoring the quality of corn planting includes: Obtain a sequence of video frames containing motion blur from corn planting; The corn sowing video frame sequence is deblurred using a pre-trained super-resolution deblurring network to obtain a clear image sequence. The super-resolution deblurring network includes a wavelet enhancement residual module, which performs wavelet domain decomposition and enhancement processing on the input features. The clear image sequence is identified by a pre-trained target detection network to obtain a corn seed target sequence. The target detection network includes a composite attention mechanism module and a multi-scale fusion module. The composite attention mechanism module combines improved coordinate attention, efficient channel attention, and parameter-free spatial attention. The multi-scale fusion module performs multi-scale feature fusion through multiple parallel convolutional branches with different dilation rates. The cumulative number of corn seed target sequences passing through a preset detection line is accumulated over time, and the relationship between the cumulative number of these sequences and a baseline is compared to determine whether a corn seed replanting event or a missed replanting event occurs.

2. The method for monitoring maize planting quality according to claim 1, characterized in that, The training process of the pre-trained super-resolution deblurring network includes: A randomized multi-stage degradation dataset is constructed, the construction process of which includes: simulating motion blur by averaging the pixels of consecutive frames from high-speed sampled clear video; and randomly applying multiple degradation stages to the generated clear-blurred image pairs, each degradation stage randomly including one or more operations among optical blur, optical chromatic aberration, sensor noise, quantization, compression artifacts, and downsampling. Using the sharp image in the sharp-blurred image pair as the supervision signal, the super-resolution deblurring network is trained on the random multi-stage degradation dataset until convergence.

3. The method for monitoring maize planting quality according to claim 1, characterized in that, The training process of the pre-trained object detection network includes: This is achieved through two-stage training, which includes: In the first training phase, small local image patches centered on the target box are cropped from the original labeled image for training to obtain the optimal weights. In the second training phase, the object detection network is initialized with the optimal weights and training continues on the original labeled images.

4. The method for monitoring corn planting quality according to claim 1, characterized in that, The specific process of performing wavelet domain decomposition and enhancement processing on the input features using the wavelet-enhanced residual module includes: Perform a two-dimensional discrete wavelet transform of the Haar wavelet basis on the input features to obtain a low-frequency sub-band and three high-frequency sub-bands in three directions, namely, a horizontal high-frequency sub-band, a vertical high-frequency sub-band, and a diagonal high-frequency sub-band. The low-frequency subband is subjected to feature transformation by 1×1 convolution and residual connection to obtain the enhanced low-frequency subband. The high-frequency subbands in the three directions are spliced ​​together in the channel dimension, and then adaptively weighted and enhanced by depthwise separable convolution and channel attention mechanism in sequence. Finally, residual connection is performed with the high-frequency subbands before splicing to obtain the enhanced high-frequency subbands. The enhanced low-frequency subband and the enhanced high-frequency subband are reconstructed into output features through inverse discrete wavelet transform.

5. The method for monitoring maize planting quality according to claim 1, characterized in that, The pre-trained super-resolution deblurring network further includes a three-stage progressive upsampling module, which receives low-resolution features output by the encoder in the super-resolution deblurring network as input, and includes: In the first upsampling stage, the low-resolution features are upsampled by a factor of two, and the low-frequency signal is enhanced by the wavelet enhancement residual module to restore the overall structure of the image. In the second upsampling stage, the features output in the first stage are upsampled by a factor of two again, and the low-frequency and high-frequency signals are simultaneously enhanced by the wavelet enhancement residual module to restore texture details. In the third upsampling stage, the high-frequency signal of the features output from the second stage is enhanced by the wavelet enhancement residual module to sharpen the image edges.

6. The method for monitoring maize planting quality according to claim 1, characterized in that, The composite attention mechanism module includes: The improved coordinate attention introduces a minimum channel number constraint during the channel compression stage and employs group normalization and the SiLU activation function. The efficient channel attention uses one-dimensional convolution with a fixed kernel size; The parameterless spatial attention introduces a numerically stable term into the variance term of the neuron energy calculation to smooth the attention weights. The composite attention mechanism module uses the improved coordinate attention as the backbone and fuses the output of the high-efficiency channel attention and the parameterless spatial attention in series with the output of the improved coordinate attention.

7. The method for monitoring maize planting quality according to claim 1, characterized in that, The multi-scale fusion module includes three parallel convolutional branches: a small receptive field branch, a medium receptive field branch, and a large receptive field branch, and the three parallel convolutional branches are allocated with asymmetric channels.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the corn planting quality monitoring method according to any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the corn planting quality monitoring method according to any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the corn planting quality monitoring method according to any one of claims 1-7.

Images