Image processing method and device

Abstract

The invention discloses an image processing method and device, relates to the technical field of image processing and intelligent agriculture, and aims to solve the problem that an existing pest control method is not suitable for crops in the early stage of pest control. The method comprises the following steps: combining a historical frame image with a current frame image to judge whether a first target in the historical frame image is a noisy point or not; and if the judgment result is no, acquiring an image which is shot at a short distance and contains the first target, and identifying the first target line in the image. According to the method, the current frame and the historical frame are combined to judge whether the spots in the historical frame are noisy points or not, so that the noisy points in the long-distance shot image can be more accurately eliminated, the accuracy of crop plant disease spot identification is improved, and a basis is provided for prevention and treatment of crop diseases and insect pests.

Description

Technical Field

[0001] The present invention relates to the technical fields of image processing and smart agriculture. Background Art

[0002] Agricultural development has entered the stage of smart agriculture. Smart agriculture uses modern information technologies such as the Internet of Things, big data, cloud computing and artificial intelligence to achieve precision, intelligence and efficiency in agricultural production management. Image processing technology, as one of the important technical means of artificial intelligence, plays an important role in the field of smart agriculture. For example, in the prevention and control of crop diseases and insect pests, drones can be used to shoot above crops, and the images taken can be processed by computers to identify some characteristics of diseases and insect pests in the images, which can play a good role in preventing and controlling crop diseases and insect pests.

[0003] When drones are taking aerial photos of crops, there is a certain height difference between the drone and the crops. This will not have much impact on shorter crops with fewer leaves, but for taller crops such as sugarcane and corn, the leaves are narrow and long, and most of them are curved. Even two adjacent leaves may have a large difference in height. Therefore, except for the top leaves, the height difference between other leaves and the drone will be relatively large. In the early stages of pests and diseases, there are often only a few very small spots on the leaves. If the drone is too high, the spot area on the image will be too small, and the characteristics of the disease spots will not be reflected. In addition, noise will appear when the camera is shooting. The time, position, shape, color, etc. of the noise are all random. Therefore, small spots often appear in the crop images taken by the camera, such as Figure 7 The spots in the middle circle are easily ignored as noise, resulting in the inability to identify the characteristics of pests and diseases during image processing.

[0004] In summary, the existing pest control methods have the following problems: they are only applicable to the stage when the pest characteristics are more obvious, but it is difficult to play a preventive role for crops in the early stage of pests and diseases. Summary of the invention

[0005] The purpose of the present invention is to solve or alleviate the above-mentioned problems in the prior art and to provide an image processing method and device.

[0006] The image processing method of the present invention includes: combining a historical frame image with a current frame image to determine whether a first target in the historical frame image is a noise point; and if the determination result is no, obtaining an image taken at a close distance and containing the first target, and identifying the first target row in the image.

[0007] Optionally, combining the historical frame image with the current frame image to determine whether the first target in the historical frame image is a noise point includes: in response to identifying the first target from the historical frame image, identifying the second target from the historical frame image, and calculating the relative position relationship between the first target and the second target in the historical frame image; identifying the first target and the second target from the current frame image, and calculating the relative position relationship between the first target and the second target in the current frame image; if the relative position relationship between the first target and the second target in the historical frame image is the same as the relative position relationship between the first target and the second target in the current frame image, then determining that the first target in the historical frame image is not a noise point.

[0008] Optionally, the first target is composed of a plurality of interconnected pixels, and the area thereof is smaller than a preset first threshold, and the pixels around the first target are all green.

[0009] Optionally, the relative position relationship includes a direction of the first target relative to the second target and a distance between the two.

[0010] Optionally, the historical frame image is a frame image previous to the current frame image.

[0011] Optionally, combining the historical frame image with the current frame image to determine whether the first target in the historical frame image is a noise point includes: sending the historical frame image and the current frame image into two channels of a dual convolutional neural network model respectively, the dual convolutional neural network model is used to classify the first target in the historical frame image, and the output category of the dual convolutional neural network model is target / noise.

[0012] Optionally, each channel of the dual convolutional neural network model includes an input layer, a convolution layer, a pooling layer, a convolution layer, a pooling layer, a convolution layer, and a pooling layer in sequence.

[0013] The present invention combines the current frame with the historical frame to determine whether the spots in the historical frame are noise points, which can more accurately exclude noise points in images taken from a long distance, improve the accuracy of identifying disease spots on crop plants, and provide a basis for the prevention and control of crop diseases and pests. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is a schematic diagram of the principle of an image processing system according to an embodiment of the present application;

[0015] Figure 2 is a schematic diagram of the structure of a computing device according to an embodiment of the present application;

[0016] Figure 3 is a flowchart of an image processing method according to an embodiment of the present application;

[0017] Figure 4 is a flowchart of step S310 according to an embodiment of the present application;

[0018] Figure 5 is a schematic diagram of the structure of a dual convolutional neural network model according to an embodiment of the present application;

[0019] Figure 6 is a schematic structural diagram of an image processing device according to an embodiment of the present application;

[0020] Figure 7 These are images of agricultural crops according to the “Background Art” section, where (a) is an image of a sugarcane plant and (b) is an image of a corn plant. DETAILED DESCRIPTION

[0021] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0022] As used herein, the singular forms "a", "an", and "the" are intended to include the plural forms as well, unless the context indicates otherwise. It should be further understood that the terms "comprise", "include" indicate the presence of the described features, operations, elements, components, items, kinds, and / or groups, but do not exclude the presence, occurrence, or addition of one or more other features, operations, elements, components, items, kinds, and / or groups. The terms "or" and "and / or" used herein are interpreted as inclusive, or mean any one or any combination. Therefore, "A, B or C" or "A, B and / or C" means "any of the following: A; B; C; A and B; A and C; B and C; A, B and C". Exceptions to this definition will only occur when the combination of elements, functions, or operations is inherently mutually exclusive in some way.

[0023] In view of the problem that the existing methods of disease and pest control in the field of smart agriculture are only applicable to the stage when the characteristics of pests and diseases are more obvious, but not to the early stages of pests and diseases, the present invention provides an image processing method that can accurately identify the characteristics of pests and diseases in the early stages of pests and diseases, providing a basis for disease and pest control work on crops.

[0024] Figure 1 FIG. 4 is a schematic diagram of an image processing system according to an embodiment of the present invention. Figure 1As shown, the image processing system includes a camera device 100 and a computing device 200, and the camera device 100 and the computing device 200 are connected via a wireless network. The camera device 100 can be installed on a drone to take photos or videos of crops driven by the drone. The computing device 200 is used to execute the image processing method of the embodiment of the present application, so as to process the image taken by the camera device 100 and give prompt information when a specific target is detected from the image. In addition, the computing device 200 can also remotely control the camera device 100 to move it according to a set trajectory.

[0025] Figure 2 FIG. 2 is a schematic diagram of a computing device 200 according to an embodiment of the present invention. It should be noted that: Figure 2 The computing device 200 shown is only an example. In practice, the computing device used to implement the image processing method of the present invention can be any type of device, and its hardware configuration can be different from the above. Figure 2 The computing device 200 shown is the same as Figure 2 The computing device 200 shown is different. In practice, the computing device used to implement the image processing method of the present invention can be Figure 2 The hardware components of the computing device 200 shown are increased or deleted, and the present invention does not limit the specific hardware configuration of the computing device.

[0026] like Figure 2 As shown, computing device 200 typically includes memory 210 , one or more processors 220 , and a communication module 230 .

[0027] Depending on the desired configuration, processor 220 may be any type of processor including, but not limited to, a microprocessor (μP), a microcontroller (μC), a digital signal processor (DSP), or any combination thereof.

[0028] Depending on the desired configuration, the memory 210 may be any type of memory including, but not limited to, volatile memory (such as RAM), non-volatile memory (such as ROM, flash memory, etc.), or any combination thereof.

[0029] The communication module 230 implements wireless communication connection between the computing device 200 and other devices.

[0030] The computing device 200 can be implemented as a server, such as a file server, a database server, an application server, and a WEB server, etc., and can also be implemented as a personal computer including a desktop computer and a notebook computer configuration. Of course, the computing device 200 can also be implemented as a part of a small-size portable (or mobile) electronic device. In an embodiment according to the present invention, the computing device 200 is configured to perform the image processing method 300 according to the present invention.

[0031] An image processing method 300 according to an embodiment of the present invention includes: combining a historical frame image with a current frame image to determine whether a first target in the historical frame image is a noise point; and if the determination result is no, obtaining an image taken at a close distance and containing the first target, and identifying the first target row in the image.

[0032] Figure 3 FIG. 3 is a flowchart of an image processing method 300 according to an embodiment of the present invention. The method 300 is executed in a computing device (such as the aforementioned computing device 200). Figure 3 As shown, the method 300 starts at step S310 .

[0033] In step S310, the historical frame image is combined with the current frame image to determine whether the first target in the historical frame image is a noise point.

[0034] In the embodiment of the present application, the area of ​​the target is represented by the number of pixels covered by the target, and the first target refers to a spot with a particularly small area. For a spot-like target, when the area is large enough, it can be determined whether it belongs to a pest feature through conventional target recognition and classification algorithms, but when the target area is too small, it is easy to be misidentified as a noise point and ignored. The embodiment of the present application sets a first threshold value, and when the area of ​​the target is identified to be smaller than the first threshold value, the target is used as the first target.

[0035] When the first target appears in a frame image, it is impossible to determine whether the first target is a noise point or a target to be identified. This step combines the situation of the first target in a frame image with the situation of the first target in the next frame image of the frame image to determine whether the first target in the frame image is a noise point.

[0036] If the judgment result is that the first target is not a noise point, the first target may be a diseased spot or an object such as dust falling on the blade, and then the process proceeds to step S320.

[0037] In step S320, an image including the first target taken at a close distance is acquired, and the first target row in the image is identified.

[0038] Since the spot (first target) is too small, it is difficult to determine whether it is a diseased spot through conventional target recognition and classification algorithms. Therefore, in this step, the drone will be controlled to lower its flight altitude to obtain a clearer first target with a larger coverage area. After reacquiring the image, conventional target recognition and classification algorithms can be used to further determine whether it is a diseased spot and the type of the diseased spot.

[0039] As a preferred embodiment of the present application, combining the historical frame image with the current frame image to determine whether the first target in the historical frame image is a noise point can be achieved through method 400.

[0040] Method 400 includes: in response to identifying a first target from a historical frame image, identifying a second target from the historical frame image, and calculating the relative position relationship between the first target and the second target in the historical frame image; identifying the first target and the second target from the current frame image, and calculating the relative position relationship between the first target and the second target in the current frame image; if the relative position relationship between the first target and the second target in the historical frame image is the same as the relative position relationship between the first target and the second target in the current frame image, then judging that the first target in the historical frame image is not a noise point.

[0041] like Figure 4 As shown, the method 400 starts at step S410 .

[0042] In step S410, in response to identifying a first target from a historical frame image, a second target is identified from the historical frame image, and a relative position relationship between the first target and the second target in the historical frame image is calculated.

[0043] In the embodiment of the present application, the historical frame refers to the previous frame image of the current frame in the video frame. The drone shoots at a lower frame rate and transmits the captured video frames to the computing device 200 in real time for image processing. In order to ensure that the entire area to be shot can be shot, the speed of the drone should not be too high.

[0044] The computing device 200 processes the received video frames in the following manner: for each frame of the image, the first target is extracted from it, and when the first target is extracted, the second target is extracted from the image.

[0045] In one implementation, on the basis of limiting the area of ​​the first target, the color of each pixel of the first target may be further limited. For example, each pixel of the first target may be white, brown, or black.

[0046] In this implementation, the first target has the following characteristics: (1) It is composed of a number of interconnected pixels; (2) The pixels are usually yellow, brown, black, or white; (3) The surrounding pixels are all green; (4) The area is smaller than a preset first threshold. It should be noted that the yellow, brown, black, or white here refers to the corresponding color system, and in the image, the first target may have only a part of the pixels in yellow / brown / black / white, while the color of the other part of the pixels is between yellow / brown / black / white and green. Taking sugarcane rust as an example, sugarcane rust mainly harms the leaves of sugarcane. In the early stage of rust, small light yellow spots appear on the leaves. In the middle stage, the light yellow spots develop into brown spots (summer spores), which are usually 2-10 mm in length and have a yellow halo on the outside. In the later stage, the spots develop into black (winter spores), and the spots are easy to connect together to form large black spots. In the early stage, there are small light yellow spots in the images taken by the drone. This spot is composed of several pixels, which may all be light yellow, or part of them may be light yellow and the other part may be between light yellow and green (the color of normal leaves). The characteristics of the spots with larger areas in the middle stage and the spots in the late stage are more obvious, and their characteristics can be identified as rust by conventional target recognition and classification algorithms, but the spots in the early stage change from small to large, and are easily misjudged as noise points when the spot area is very small. In the embodiment of the present application, it is assumed that the first threshold is set to 9 pixels. If a spot with an area less than 9 pixels appears in the image, and a part of the pixels are light yellow and another part of the pixels are yellow-green, then the spot is used as the first target.

[0047] The computing device 200 identifies the first target for each frame of the received video frame. In a frame, the first target may or may not exist. If it exists, the number of first targets may be one, two or more. If the first target is identified in a frame, the second target is identified in the frame. In the embodiment of the present application, the second target refers to a complete leaf or a partial leaf, and the recognition of the second target can be implemented using a convolutional neural network.

[0048] In one implementation, the second target is a complete leaf. In a frame of image, each first target has a corresponding second target, and the second target corresponding to a first target should be a complete leaf, and the first target is located inside the second target, that is, the pixels covered by the second target include the pixels covered by the first target.

[0049] In another implementation, the second target is a part of the blade. It should be noted that the part here should include a partial outline of the blade, and the distance between the partial outline and the first target should be as small as possible, so as to ensure that the second target can appear in several adjacent frame images at the same time. Specifically, if multiple incomplete blade outlines are identified in the blade, the one closest to the first target can be selected as the second target. Furthermore, the direction can also be limited. For example, the incomplete blade outline located on the right side of the first target and closest to the first target is selected as the second target corresponding to the first target. Regardless of the method of selecting the second target, the purpose is to ensure as much as possible that the second target matching the same first target in two adjacent frame images is the same second target, thereby improving the accuracy of the judgment.

[0050] After the first target and the second target corresponding to the first target are identified in the same frame image, the relative position relationship between the first target and the second target is calculated.

[0051] The relative position relationship includes the direction of one target relative to another target and the distance between the two. Specifically, it can be calculated by the following steps:

[0052] First, determine the center point of the first target and the center point of the second target, and record the position coordinates of the two center points in the image pixel matrix. It should be noted that the center point only contains one pixel, and the center points of the two targets are respectively called the first center point and the second center point;

[0053] Then, the direction of the first center point relative to the second center point and the distance between the first center point and the second center point are calculated according to the position coordinates of the center point in the pixel matrix. The direction of the first center point relative to the second center point can be represented by the angle between the line connecting the first center point to the second center point and the reference direction. The reference direction can be set to the direction of the pixel matrix row, and rightward is positive.

[0054] Next, in step S420, the first target and the second target are identified from the current frame image, and the relative position relationship between the first target and the second target in the current frame image is calculated.

[0055] The method of identifying the first target and the second target in this step is the same as that in the previous step.

[0056] Next, in step S430, if the relative position relationship between the first target and the second target in the historical frame image is the same as the relative position relationship between the first target and the second target in the current frame image, it is determined that the first target in the historical frame image is not a noise point.

[0057] The time and position of the noise point are random, but the disease spots on the leaves of the plant always exist. If there is a pair of first and second targets with the same relative position relationship in two consecutive frames, it can be considered that the first target in the previous frame is not a noise point. On the contrary, assuming that the first target and the second target are identified in two consecutive frames, but the relative position relationship between the first and the second target in the two frames is different, it can be considered that the first target in the previous frame is a noise point, and whether the first target in the next frame is a noise point needs to be confirmed based on subsequent images.

[0058] It should be noted that the same relative position relationship between the first target and the second target in the two frames means that the difference in the distance and direction of the first target relative to the second target in the two frames is within a reasonable range. For example, if the distance error δ is set to 3 pixels and the direction error Δ is set to 0.5°, the distance and direction of the first target relative to the second target in the first frame are d1 and α1 respectively, and the distance and direction of the first target relative to the second target in the second frame are d2 and α2 respectively. | d1-d 2| ≤δ, and | α1-α 2| ≤Δ, then the first target in the two images is considered to be the same target, and the first target in the current frame image is not a noise point.

[0059] As a preferred embodiment of the present application, step S310 combines the historical frame image with the current frame image to determine whether the first target in the historical frame image is a noise point. It can also be achieved by the following method: the historical frame image and the current frame image are respectively sent to two channels of a dual convolutional neural network model, and the dual convolutional neural network model is used to classify the first target in the historical frame image, and the output category of the dual convolutional neural network model is target / noise.

[0060] The current frame image and the frame image before the current frame image are respectively sent to two channels of the dual convolutional neural network model, and the dual convolutional neural network model is used to classify the first target in the frame image before the current frame.

[0061] In the embodiment of the present application, the structure of the dual convolutional neural network model is as follows Figure 5 As shown in the figure, the model contains two channels, each of which contains input layer, convolution layer, pooling layer, convolution layer, pooling layer, convolution layer, pooling layer, and the last pooling layer is followed by a fusion layer. The fusion layer fuses the outputs of the convolution modules of the two channels, and the obtained data is used as the input of the fully connected layer. The fully connected layer is followed by the output layer. The convolution kernel sizes of the three convolution layers of each convolution channel are 5×5, 3×3, and 3×3, respectively, with a step size of 1; the pooling layer window size is 2×2, with a step size of 2, and the maximum pooling method is adopted; the activation function adopts the ReLU function.

[0062] The dual convolutional neural network model can classify spot-like targets in the input image, and the classification results are noise or targets.

[0063] If the same spot appears in two images input to the dual convolutional neural network model, the spot is likely to be the target. If the spot exists in the previous image but not in the next image, the spot in the previous image is a noise point.

[0064] It should be noted that each image of the video frame will be input into the dual convolutional neural network model twice, the first time as the current frame image and the second time as the historical frame.

[0065] Aerial images of various crops are divided into training sets and test sets. It should be noted that no matter in the training set or the data set, the images should appear in pairs, and each pair of images comes from two adjacent frames in the aerial video. For any crop, the method of collecting training data and test data is as follows:

[0066] Aerial photography was performed on a planting area that was not infected with any pests or diseases and had clean leaves to obtain the first data set, in which all data were labeled as "noise points";

[0067] Small granular objects are sprayed on the leaves of the plants in the area, and the area is photographed again by aerial means to obtain a second data set, in which all data are labeled as "target";

[0068] The planting area of ​​the crop variety that is in the early stage of infection by various pests and diseases is selected, and the area is photographed by aerial photography to obtain the third data set. The label of all data in this data set is "target".

[0069] Collecting data in the above manner can avoid the time and effort consumed by manual labeling, and can also avoid the inaccuracy caused by automatic labeling.

[0070] The first data set is divided into a first training data set and a first test data set according to a certain ratio (for example, 8:2), the second data set is divided into a second training data set and a second test data set according to the same ratio, and the third data set is divided into a third training data set and a third test data set according to the same ratio. The data of the first training data set, the second training data set and the third training data set are merged as the training data set, and the data of the first test data set, the second test data set and the third test data set are merged as the test data set.

[0071] The images in the data set should appear in pairs. Two adjacent frames of aerial data are considered as a pair of images, and the aerial data is divided into multiple pairs of data. You can also take a pair of images from the aerial video at regular intervals, for example, every 0.5 seconds. Since the drone flies at a low speed, the same blade will appear in multiple consecutive frames in an aerial video. Taking a pair of images at regular intervals can eliminate most of the redundant data.

[0072] The image pairs of the training data set are used as model training data, and the image pairs of the test data set are used as model test data to ensure that data from the same sample only appears in the training set or the test set during the training-testing process. The training-testing process of the model can be repeated multiple times, and the average of the multiple test results is used as the final recognition result of the dual convolutional neural network model.

[0073] As a preferred embodiment of the present application, step S310 combines the historical frame image with the current frame image to determine whether the first target in the historical frame image is a noise point, which can also be achieved by the following method:

[0074] First, the drone is controlled to fly along a preset trajectory. The trajectory is divided into multiple segments, each of which is a straight line and the length of each segment does not exceed 1 meter. For each segment, the drone flies from the starting point to the end point, and then returns from the end point to the starting point. The drone takes pictures of crops during the flight, with a frequency of 5-10Hz. The drone is equipped with a positioning device, and each image sent back by the drone is accompanied by the shooting time and location information. In this way, for each location information, there are two images with different shooting times.

[0075] Next, the computing device 200 identifies the first target for each received image.

[0076] Usually, there are multiple first targets in an image, and the center point of each first target is taken as the position of the first target, and the position coordinates of each first target in the pixel matrix are recorded.

[0077] The relative position relationship between the first targets in an image can be expressed as follows:

[0078] Next, two images with the same position information (hereinafter referred to as the first image and the second image) are compared.

[0079] If the relative position relationship between several first targets in the first image is the same as the relative position relationship between several first targets in the second image, it means that these first targets are "targets" rather than noise points. At this time, the drone can be controlled to return to the shooting position of the first image and the second image, and the drone can be lowered to reacquire the close-range image, and then the close-range image can be identified and classified.

[0080] If the relative position relationship between each first target in the first image is completely different from the relative position relationship between each first target in the second image, it means that the first targets in the first image and the second image are all noise points and can be ignored.

[0081] The method of determining whether the first target is a noise point in this embodiment does not need to adopt a convolutional neural network model, but only needs to perform simple mathematical calculations based on the horizontal and vertical coordinates, thereby avoiding the training of the convolutional neural network model and the creation of training data sets and test data sets. The computing device 200 has less computational complexity, thereby improving the real-time performance of the image processing process.

[0082] The method 300 of the embodiment of the present application combines the current frame with the historical frame to determine whether the spots in the historical frame are noise points. This can more accurately exclude noise points in images taken from a long distance, improve the accuracy of identifying disease spots on crop plants, and provide a basis for the prevention and control of crop diseases and pests.

[0083] The embodiment of the present invention further provides an image processing device 600, which can execute the processing of each step of the image processing method 300 described above. Figure 6 The above-mentioned image processing apparatus 600 will be described.

[0084] like Figure 6 As shown, the device 600 includes a judgment unit 610 and a target recognition unit 620 .

[0085] The determination unit 610 is adapted to combine the historical frame image with the current frame image to determine whether the first target in the historical frame image is a noise point.

[0086] The target recognition unit 620 is adapted to, if the judgment result is no, obtain an image taken at a close distance and containing the first target, and recognize the first target in the image.

[0087] As a preferred embodiment of the present application, the determination unit 610 includes:

[0088] a first recognition unit adapted to recognize a second target from the historical frame image in response to recognizing a first target from the historical frame image, and calculate a relative position relationship between the first target and the second target in the historical frame image;

[0089] A second recognition unit, adapted to recognize the first target and the second target from the current frame image, and calculate the relative position relationship between the first target and the second target in the current frame image;

[0090] The relative position relationship judgment unit is adapted to judge that the first target in the historical frame image is not a noise point if the relative position relationship between the first target and the second target in the historical frame image is the same as the relative position relationship between the first target and the second target in the current frame image.

[0091] As a preferred embodiment of the present application, the first target is composed of a plurality of interconnected pixels, and the area thereof is smaller than a preset first threshold value, and the pixels around the first target are all green.

[0092] As a preferred embodiment of the present application, the relative position relationship includes the direction of the first target relative to the second target and the distance between the two.

[0093] As a preferred embodiment of the present application, the historical frame image is the previous frame image of the current frame image.

[0094] As a preferred embodiment of the present application, the judging unit judges whether the first target in the historical frame image is a noise point in the following manner:

[0095] The historical frame image and the current frame image are respectively sent to two channels of a dual convolutional neural network model, and the dual convolutional neural network model is used to classify the first target in the historical frame image, and the output category of the dual convolutional neural network model is target / noise.

[0096] As a preferred embodiment of the present application, each channel of the dual convolutional neural network model includes an input layer, a convolutional layer, a pooling layer, a convolutional layer, a pooling layer, a convolutional layer, and a pooling layer in sequence.

[0097] The image processing device 600 of the embodiment of the present application has the same principles and effects as the image processing method 300, which will not be described in detail herein.

[0098] The various techniques described herein may be implemented in combination with hardware or software, or a combination thereof. Thus, the method and apparatus of the present invention, or some aspects or parts of the method and apparatus of the present invention may be in the form of program codes (i.e., instructions) embedded in a tangible medium, such as a removable hard disk, a USB flash drive, a floppy disk, a CD-ROM, or any other machine-readable storage medium, wherein when the program is loaded into a machine such as a computer and executed by the machine, the machine becomes a device for practicing the present invention.

Claims

1. An image processing method, characterized in that: include: The historical frame image is combined with the current frame image to determine whether the first target in the historical frame image is a noise point; as well as If the judgment result is no, an image taken at a close distance and containing the first target is obtained, and the first target in the image is identified.

2. The method according to claim 1, characterized in that The combining of the historical frame image with the current frame image to determine whether the first target in the historical frame image is a noise point includes: In response to identifying a first target from the historical frame image, identifying a second target from the historical frame image, and calculating a relative positional relationship between the first target and the second target in the historical frame image; Identify a first target and a second target from a current frame image, and calculate a relative position relationship between the first target and the second target in the current frame image; If the relative position relationship between the first target and the second target in the historical frame image is the same as the relative position relationship between the first target and the second target in the current frame image, then it is determined that the first target in the historical frame image is not a noise point.

3. The method according to claim 2, characterized in that The first target is composed of a plurality of interconnected pixels, and the area thereof is smaller than a preset first threshold value, and the pixels around the first target are all green.

4. The method according to claim 2, characterized in that The relative position relationship includes the direction of the first target relative to the second target and the distance between the two.

5. The method according to claim 2, characterized in that The historical frame image is the frame image before the current frame image.

6. The method according to claim 1, characterized in that The combining of the historical frame image with the current frame image to determine whether the first target in the historical frame image is a noise point includes: The historical frame image and the current frame image are respectively sent to two channels of a dual convolutional neural network model, and the dual convolutional neural network model is used to classify the first target in the historical frame image, and the output category of the dual convolutional neural network model is target / noise.

7. The method according to claim 6, characterized in that Each channel of the dual convolutional neural network model includes an input layer, a convolution layer, a pooling layer, a convolution layer, a pooling layer, a convolution layer, and a pooling layer in sequence.

8. An image processing device, characterized in that: include: a judgment unit, adapted to combine the historical frame image with the current frame image to judge whether the first target in the historical frame image is a noise point; as well as The target recognition unit is adapted to, if the judgment result is no, obtain an image taken at a close distance and containing the first target, and recognize the first target in the image.

9. A computing device comprising: at least one processor and a memory storing program instructions; When the program instructions are read and executed by the processor, the computing device executes the image processing method according to any one of claims 1 to 7.

10. A readable storage medium storing program instructions, when the program instructions are read and executed by a computing device, the computing device executes the image processing method according to any one of claims 1 to 7.

Images