A digital measurement system for the root length of potatoes and its measurement method

Abstract

The present invention belongs to the field of instruments, and particularly relates to a digital measurement system for the root length of potatoes and a measurement method thereof. The measurement system includes an image acquisition device and a data processing device. The image acquisition device acquires a sample image of the potato roots to be measured; the data processing device is communicatively connected to the image acquisition device and is used to control the image acquisition device. The data processing device includes a data acquisition module, an image segmentation module, a coefficient query module, and a length calculation module. The image segmentation module contains an image segmentation network based on DeepLabV3+ semantic segmentation, and a pixel conversion coefficient table is preset in the coefficient query module. The length calculation module first performs refinement processing on the image segmentation result, and then calculates the true root length according to the root length represented by the pixels corresponding to the potato roots and the pixel conversion coefficient. The present invention solves the problems of difficult manual measurement of the root length of shallow-rooted crops such as potatoes, long task duration, large workload, cumbersome process, and low efficiency.

Description

Technical Field

[0001] The present invention belongs to the field of instruments, and particularly relates to a digital measurement system for potato root length and a measurement method thereof. Background Art

[0002] The root system is the main medium connecting plants and the soil environment, and is the link for plants to establish mutual relationships with the rhizosphere environment. After absorbing water and various mineral nutrients from the soil, the root system transports them to the above-ground part of the plant for its use, and at the same time fixes the plant in the soil. Root architecture is the spatial structure and distribution of crop root systems in the growth medium, which not only determines the strength of the plant fixing ability, but also is closely related to the plant's ability to absorb and utilize water and nutrients in the soil. The root system can not only sense changes in environmental factors, but also improve the plant's absorption and utilization of scarce resources (such as water and nutrients) through changes in its morphological characteristics (such as root length, root weight, and root volume), spatial distribution (such as root angle, rooting depth, etc.), anatomical structure (such as aerenchyma), and metabolic activity (such as the absorption rate and respiration rate of water and fertilizer), avoid or reduce damage to the plant caused by stress, and maintain a relatively high economic yield and biological yield as much as possible.

[0003] In the soil profile, root architecture is jointly determined by the genetic structure of the root system and soil environmental factors. The study of root architecture is very important for agricultural production. Due to the extremely complex soil composition, the distribution of nutrient resources is often uneven, and its effectiveness has great variability in both time and space. The plant root system is always faced with the heterogeneity of soil nutrient resource supply. Therefore, studying the growth and distribution characteristics of roots in the soil and analyzing the plant's ability to absorb water and nutrients are important indicators for evaluating crop variety characteristics and the performance of products such as pesticides and fertilizers. It is crucial for improving the scientific and technological R & D level of germplasm and agricultural materials and realizing agricultural modernization.

[0004] One of the tasks of root research is to measure the length of roots. Different types of crops have different root types. Taking potatoes as an example, the root system includes a main root, lateral roots, and abundant fibrous roots. When measuring the root length, it is usually necessary to count all the roots. Most of the conventional crop root length measurement tasks are completed manually. Technicians need to divide the entire root system into discrete roots one by one, and then measure and count each root in turn.

[0005] In a research project on potato water-saving and high-yield completed by the inventor of this case, a large amount of data related to potato root length was required. Different crops have different numbers of roots, and the measurement difficulty varies greatly. For shallow-rooted crops such as potatoes, because the number of main roots is small, while the number of lateral roots and fibrous roots is very abundant, it is extremely cumbersome to measure the potato root length. Summary of the Invention

[0006] In order to solve the problems of difficult manual measurement of the root length of shallow-rooted crops such as potatoes, long task duration, large workload, cumbersome operation, and low efficiency, the present invention provides a digital measurement system and a measurement method for the root length of potatoes.

[0007] The present invention is implemented by the following technical solutions:

[0008] A digital measurement system for the root length of potatoes, which includes an image acquisition device and a data processing device. The image acquisition device is used to obtain a sample image of the potato root system sample to be measured; the data processing device is used to measure the length of the potato root system sample according to the sample image.

[0009] The image acquisition device includes a sample stage, a bracket, a camera, and a fill light. The sample stage includes a horizontal operation platform. The center of the operation platform contains a detachable transparent sample placement board. The upper surface of the sample placement board is the sample area for placing the potato root system sample to be measured. The inside of the sample placement board contains a sandwich space for placing a solid-color background board. The camera and the fill light are fixedly connected above the sample stage through the bracket. The camera is used to collect a sample image of the target to be measured on the background board along the orthographic projection direction, and the fill light is used to provide uniform ambient light during the camera shooting process.

[0010] The data processing device is communicatively connected to the image acquisition device and is used to control the image acquisition device. The data processing device includes a data acquisition module, an image segmentation module, a coefficient query module, and a length calculation module. The data acquisition module is used to obtain the sample image taken by the camera and extract the shooting information corresponding to the sample image during shooting. The image segmentation module contains an image segmentation network for identifying the root system sample and segmenting the identified root system sample. The image segmentation network is trained based on an improved DeepLabV3+ network model. The image segmentation module is used to perform feature recognition and image segmentation on the input sample image, and then output a segmented target image that only contains the root system part. The coefficient query module presets a pixel conversion coefficient table for representing the mapping relationship between pixels and the true scale in the sample image obtained by the camera in different shooting states. And it is used to query the corresponding pixel conversion coefficient B according to the shooting information of the sample image extracted by the data acquisition module. The length calculation module is used to first optimize the pixels of the target image through the Hilditch thinning algorithm and calculate the root length P represented by the pixels of the potato root system in the sample image L ; Then, in combination with the pixel conversion coefficient B obtained by the coefficient query model, use the formula R L = B·P L to calculate the true root length R L .

[0011] As a further improvement of the present invention, the image segmentation network selects the DeepLabV3+ network model with an encoder-decoder structure as the basic model, and replaces the backbone network Xception in the basic model with the MobileNetV2 network. In the network model, the CARAFE upsampling module is used to replace the Upsample by 4 upsampling module in the basic network, and the high-level features output by the encoder and the non-output results of the decoder are upsampled. And a CBAM attention mechanism module is added respectively after the low-level features output by the DCNN and the feature connection layer of the encoder.

[0012] As a further improvement of the present invention, the working process of the image segmentation network is as follows: First, the input sample image extracts image features through the MobileNetV2 backbone network in the encoder part. Secondly, it enters the atrous spatial pyramid pooling module to obtain image spatial feature information and transmits it to the CBAM module. Then, the input feature layer of the CBAM module undergoes a pooling operation through the channel attention module to obtain the weights of each channel of the input feature layer, and applies them to the spatial attention module. The spatial attention module takes the maximum and average values on the channels of each feature point, and then through the same operation as the channel attention, obtains the weights of each feature point of the input feature layer. Then multiply the weights with the original input feature layer and obtain deep features containing multi-scale context information after convolution processing. Next, in the decoder part, the extracted original features are sent into the CBAM module, and shallow features containing multi-scale context information are obtained after the same processing. Finally, feature extraction is performed on the fused image through operations such as CARAFE upsampling and convolution to accurately segment the potato root system part in the input sample image.

[0013] As a further improvement of the present invention, in the image segmentation module, the training method of the image segmentation network is as follows:

[0014] S1: Select the true root system samples of the potato plants to be measured, and after preprocessing the root system samples, evenly lay them on the sample placement board of the image acquisition device.

[0015] S2: Replace different root system samples respectively, and adjust the morphology, position of the root system samples, and the shooting parameters of the camera in the image acquisition device to obtain a large number of different root system sample images, and constitute the required sample data set.

[0016] S3: Use the data set augmentation method to expand the number of sample images in the sample data set, and then divide the sample data set into a training set and a validation set.

[0017] S4: Labelme image annotation tool is used to annotate the characteristic parts of potato root systems in the training set; the annotated training set includes two semantic classifications, namely the foreground category and the background category.

[0018] S5: The designed image segmentation network is trained using the sample images in the pre-annotated training set: the training process is as follows:

[0019] S51: The following mean intersection over union is used as an index to evaluate the performance of the segmentation model, and the calculation formula is as follows:

[0020]

[0021] In the above formula, k is the total number of semantic categories; TP represents the accuracy rate of the model predicting root pixel values, TN represents the accuracy rate of the model predicting non-root pixel values; FP represents the false positive rate of the model; FN represents the false negative rate of the model predicted by the model;

[0022] S52: The training parameter settings during the preset training process are as follows: using a dynamic learning rate, setting the initial learning rate value to 0.1, setting the batch size to 16, and training for 60 epochs.

[0023] S53: The image segmentation network is trained using the sample images in the training set after adding annotations; among them, the order of the sample images in the dataset is randomly rearranged before each round of training.

[0024] S54: After reaching the preset number of epochs, the training process of the network model is ended, and the trained network model is verified through the validation set; the model parameters of the image segmentation network after training are retained.

[0025] As a further improvement of the present invention, the dataset augmentation method adopted in step S3 includes image rotation processing, image mirror processing, and image enhancement processing; the methods of image enhancement processing include brightness change, sharpness change, and image blurring.

[0026] As a further improvement of the present invention, in the coefficient query module, the generation method of the pixel conversion coefficient table is as follows:

[0027] (1) Obtain an image of a real calibrated ruler placed on the operating platform through an image acquisition device.

[0028] (2) Perform median filtering on the image of the ruler, and use an iterative segmentation method to segment the region of interest, find the centimeter marks on the ruler, and use morphological transformation to remove the millimeter marks.

[0029] (3) Obtain the horizontal projection of the centimeter marks on the ruler through Radon transform, and finally determine the intermediate interval value D between adjacent centimeter marks on this projection i。

[0030] (4) Calculate the conversion factor between "pixels - millimeters", that is, the pixel conversion coefficient B, according to the corresponding relationship between the intermediate interval value D between adjacent centimeter marks i and the pixels in the image. The calculation formula is as follows:

[0031]

[0032] In the above formula, P d is the pixel value between adjacent centimeter marks; D i is the intermediate interval value between adjacent centimeter marks; the unit of the pixel conversion coefficient B is: pixels / mm.

[0033] (5) Adjust the shooting state of the image acquisition device, and repeat steps (1) to (4) to determine the pixel conversion coefficients corresponding to different shooting states, and record the mapping relationship between the shooting information and the pixel conversion coefficients to obtain the required pixel conversion coefficient table.

[0034] As a further improvement of the present invention, the bracket in the image acquisition device adopts an electric lifting bracket; the supplementary light adopts a multi - light source shadowless lamp system. The supplementary light is located directly above the sample placement plate, and each light source is distributed in a ring shape. The camera in the image acquisition device is located in the center of the light source. The illumination area of the supplementary light is the sample area in the lower sample stage, so that the brightness of different positions in the sample area is kept uniform when imaging in the camera. The light source in the supplementary light adopts a cold light lamp with a color temperature close to natural light.

[0035] As a further improvement of the present invention, the image acquisition device further includes a foldable light - shielding cover. The light - shielding cover is in a sleeve shape in the unfolded state and covers the outer periphery of the sample area of the sample placement plate; the light - shielding cover is detachably connected to the operation platform, and a soft light film is provided on the inner wall of the light - shielding cover.

[0036] As a further improvement of the present invention, the image processing device includes a display screen and an input device; the image processing device also runs a human - computer interaction system. The digital measurement system interacts with the operator through the human - computer interaction system to enable a technician to manipulate the operation process of the digital measurement system.

[0037] The present invention also includes a digital measurement method for the length of potato roots. The aforementioned digital measurement system for potato roots is designed based on this digital measurement method for the length of potato roots and is used to digitally measure the total length of potato root samples. The digital measurement method for the length of potato roots provided by the present invention includes the following steps:

[0038] S01A: Set up an image acquisition device for obtaining a sample image of the target to be measured in the orthographic projection direction.

[0039] S02A: Design an image segmentation network based on deep learning technology to extract the object to be measured from the collected sample images.

[0040] S03A: Use an image acquisition device to capture test images of a real ruler with different shooting parameters, and calculate the pixel conversion coefficient of the image acquisition device in different shooting states according to the object-image ratio in the test images.

[0041] The pixel conversion coefficient is the conversion factor between the pixel representation length and the real length in the sample images captured by the image acquisition device under different shooting parameters.

[0042] S04A: Segment and flatten the sample potato root system to be measured, place it on the image acquisition device, adjust the shooting parameters of the image acquisition device to obtain a sample image with the best image quality, and determine the pixel conversion coefficient corresponding to the current shooting parameters.

[0043] S05A: Input the sample image into the image segmentation network, and the image segmentation network performs feature extraction and image segmentation on the sample image to obtain the local root image after image segmentation.

[0044] S06A: Use the Hilditch thinning algorithm to thin the obtained local root image, remove the isolated interfering pixels in the local root image, and output the pixel values corresponding to the local root image.

[0045] S07A: Obtain the root length P represented by the pixels corresponding to all root hairs in the optimized local root image L , and calculate the real root length R through the following formula L :

[0046] R L = B·P L

[0047] In the above formula, B is the known pixel conversion coefficient between the pixel length and the real length of the object in the sample image.

[0048] The technical solution provided by the present invention has the following beneficial effects:

[0049] In view of the many drawbacks of manual measurement of potato root length, the present invention uses computer vision technology to design a complete set of digital measurement solutions for potato root length, including corresponding digital measurement methods and digital measurement systems. In the digital measurement method, a new image segmentation network dedicated to potato root feature recognition and extraction is mainly designed. The network model has the characteristics of strong robustness, short model training time, and few parameters, and shows advantages such as high accuracy and fast segmentation rate in the sample image segmentation of crops with more fibrous roots such as potatoes. At the same time, the present invention also accurately generates the mapping relationship between the pixel length of the target to be measured in the sample image and the real scale under the shooting mode of different sample images through the design and testing of the image acquisition equipment. In the scheme of the present invention, the total length of all root samples contained in the sample image can be quickly and accurately calculated based on the local image of the potato root segmented by the image segmentation network and the pixel conversion coefficient of the corresponding sample image.

[0050] The digital measurement method provided by the present invention can simultaneously measure the total length of multiple roots through one measurement task. In addition, no operator is required to perform operation or calculation processing during the measurement process; it is very convenient and efficient. In addition, after retraining and simple model parameter adjustment, the digital measurement system of potato root length of the device of the present invention can be used for the measurement of different types of potato roots, with high measurement accuracy and good adaptability. It can be widely used in the measurement of the length of various crop samples in the field of agricultural research, and has strong practicality.

[0051] In the digital measurement system provided by the present invention, the image acquisition device for acquiring the sample image of the target to be measured is also optimized and designed. The newly designed image acquisition device can accurately obtain the image of the target to be measured in the positive projection direction. And through the clever structural design and optimized light control, the quality of the sample image taken is significantly improved, the brightness of the sample image is uniform, the contrast between the target to be measured and the background is strong, and the image edge is not distorted; thus, the measurement accuracy of the digital measurement solution can be significantly improved, reaching a level comparable to or even exceeding the manual measurement accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0053] Figure 1 This is a flowchart of the steps of a digital measurement method for potato root length provided in Example 1 of the present invention.

[0054] Figure 2 This is a schematic diagram of the structure of a digital measurement system for potato root length provided in Example 2 of the present invention.

[0055] Figure 3 This is the principle framework diagram of each functional model in the data processing in Embodiment 2 of the present invention.

[0056] Figure 4 This is the network architecture diagram of the classic DeepLabV3+ network model.

[0057] Figure 5 This is the functional principle diagram of the atrous spatial pyramid pooling module in the DeepLabV3+ network.

[0058] Figure 6 This is the network architecture diagram of the image segmentation network improved based on the DeepLabV3+ network in Embodiment 2 of the present invention.

[0059] Figure 7 This is the functional principle diagram of the CBAM attention mechanism functional module adopted in the improved image segmentation network in Embodiment 2 of the present invention.

[0060] Figure 8 This is the step schematic diagram of the network training process of the image segmentation network in Embodiment 2 of the present invention.

[0061] Figure 9 This is a case diagram of mirroring and rotating the original image in the sample dataset.

[0062] Figure 10 This is a case diagram of image enhancement processing on the original image in the sample dataset.

[0063] Figure 11 This is the flowchart of the pixel conversion coefficient table generation method.

[0064] Figure 12 This is a simple device picture of the image acquisition device built in the performance test stage.

[0065] Figure 13 This is the login interface of the computer system in the data processing device built in the performance test stage.

[0066] Figure 14 This is the calculation interface of the computer system in the data processing device built in the performance test stage.

[0067] Figure 15 This is a schematic diagram of using the Labelme image labeling tool to label the sample image in the performance test stage.

[0068] Figure 16 This is the loss value curve of the image segmentation network in the training stage in the performance test stage.

[0069] Figure 17It is a case diagram of a sample image containing only the main root system during the performance test stage after being processed by an image segmentation network.

[0070] Figure 18 It is a case diagram of a sample image of a complex root system containing the main root and lateral roots during the performance test stage after being processed by an image segmentation network.

[0071] Figure 19 It is a case diagram of using the Hilditch thinning algorithm to process the segmented local root system image during the performance test stage.

[0072] Figure 20 It is the fitting curve of the measurement accuracy between the digital measurement scheme and the manual measurement scheme in Embodiment 2 of the present invention. Detailed implementation manners

[0073] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0074] Embodiment 1

[0075] This embodiment provides a digital measurement method for the length of potato root systems, which is used to digitally measure the total length of potato root system samples. This measurement method is different from the conventional manual measurement scheme. When measuring, technicians only need to clean, dry the sample to be measured and then perform segmentation processing, and then collect the sample image of the root system sample in a flat state, and then perform data processing on the sample image according to a preset processing flow to obtain the length measurement result of the root system sample.

[0076] One of the most prominent advantages of this measurement scheme is that this method can simultaneously measure a large number of root system samples synchronously. When processing crop samples with rich root systems, only one photo needs to be taken to measure the length measurement result of the crop sample. Therefore, it can greatly improve the measurement efficiency of potato root system samples and reduce the workload and labor intensity of technicians. This scheme is a fast and efficient automated measurement scheme.

[0077] Specifically, as Figure 1 shown, the digital measurement method for the length of potato root systems provided in this embodiment includes the following steps:

[0078] S01A: Set up an image acquisition device for obtaining a sample image of the target to be measured in the orthographic projection direction.

[0079] S02A: Design an image segmentation network based on deep learning technology for extracting the target object to be measured from the collected sample image.

[0080] S03A: Obtain test images by photographing a real ruler with an image acquisition device at different shooting parameters, and calculate the pixel conversion coefficient of the image acquisition device in different shooting states according to the object-image ratio in the test images.

[0081] The pixel conversion coefficient is the conversion factor between the pixel representation length and the real length in the sample images captured by the image acquisition device under different shooting parameters.

[0082] S04A: Segment and flatten the sample of the potato root system to be measured, place it on the image acquisition device, adjust the shooting parameters of the image acquisition device to obtain a sample image with the best image quality, and determine the pixel conversion coefficient corresponding to the current shooting parameters.

[0083] S05A: Input the sample image into an image segmentation network, and the image segmentation network performs feature extraction and image segmentation on the sample image to obtain the local root system image after image segmentation.

[0084] S06A: Use the Hilditch thinning algorithm to thin the obtained local root system image, remove the isolated interfering pixels in the local root system image, and output the pixel values corresponding to the local root system image.

[0085] S07A: Obtain the root system length P represented by pixels corresponding to all root hairs in the optimized local root system image L , and calculate the real root system length R through the following formula L :

[0086] R L = B·P L

[0087] In the above formula, B is the known pixel conversion coefficient between the pixel length and the real length of the target in the sample image.

[0088] Embodiment 2

[0089] This embodiment provides a digital measurement system for the length of potato root systems. This type of digital measurement system is a corresponding measurement device designed based on the digital measurement method in Embodiment 1. As Figure 2 shown, the digital measurement system includes an image acquisition device and a data processing device. The image acquisition device is used to obtain a sample image of the potato root system sample to be measured; the data processing device is used to measure the length of the potato root system sample according to the sample image.

[0090] The image acquisition device includes a sample stage, a bracket, a camera, and a fill light. The sample stage includes a horizontal operation platform. The center of the operation platform contains a detachable transparent sample placement plate. The upper surface of the sample placement plate is the sample area for placing potato root samples to be measured. Inside the sample placement plate, there is a sandwich space for placing a solid-color background plate. The camera and the fill light are fixedly connected above the sample stage through the bracket. The camera is used to collect sample images of the target to be measured on the background plate along the orthographic projection direction, and the fill light is used to provide uniform ambient light during the camera shooting process.

[0091] The data processing device is communicatively connected to the image acquisition device and is used to control the image acquisition device. In an actual product solution, the data processing device can be used as a host computer of the image acquisition device. The host computer receives the sample data collected by the image acquisition device and processes the sample data by running a specific calculation program to output the corresponding measurement results. According to the functional module division, the functional model architecture of the data processing device is roughly as Figure 3 shown. The data processing device includes a data acquisition module, an image segmentation module, a coefficient query module, and a length calculation module.

[0092] In the data processing device designed in this embodiment, the data acquisition module is used to obtain the sample images taken by the camera and extract the corresponding shooting information when the sample images are taken. The image segmentation module contains an image segmentation network for identifying root samples and segmenting the identified root samples. The image segmentation network is trained based on an improved DeepLabV3+ network model. The image segmentation module is used to perform feature recognition and image segmentation on the input sample images, and then output a target image that only contains the root part after segmentation. The coefficient query module is preset with a pixel conversion coefficient table for characterizing the mapping relationship between pixels and the true scale in the sample images obtained by the camera under different shooting states. And it is used to query the corresponding pixel conversion coefficient B according to the shooting information of the sample images extracted by the data acquisition module. The length calculation module is used to first optimize the pixels of the target image through the Hilditch thinning algorithm and calculate the root length P represented by the pixels of the potato roots in the sample images L ; then, in combination with the pixel conversion coefficient B obtained by the coefficient query model, use the formula R L = B·P L to calculate the true root length R L .

[0093] In the technical solution provided in this embodiment, DeepLabV3+ is a semantic segmentation model based on a deep convolutional neural network, and this network can achieve pixel-level segmentation of images. The network architecture of the traditional DeepLabV3+ network model is roughly as Figure 4As shown, compared with other semantic segmentation networks, an encoder-decoder structure is introduced in the DeepLabV3+ network model. At the same time, the Atrous Spatial Pyramid Pooling (ASPP) method is integrated with the encoder-decoder structure to fully exploit context multi-scale feature information, extract more pixel information, and better capture object boundaries.

[0094] The overall architecture of the DeepLabv3+ semantic segmentation network is divided into two parts: an encoder and a decoder. The network first passes through the encoder part. In this part, the input sample image first undergoes feature extraction through the backbone network to obtain deep features and shallow features. Then, it passes through the Atrous Spatial Pyramid Pooling (ASPP) module, where atrous convolutions with different dilation rates are used to extract semantic information from the deep feature map. After that, a 1×1 convolutional layer is used to adjust the number of channels, obtaining the deep semantic feature information of the sample image, which is then transmitted to the decoder module for processing.

[0095] In the decoder part, shallow features are first extracted from the backbone network, and the shallow feature map is convolved through a 1×1 convolution to reduce its number of channels. Then, the deep features output by the encoder are upsampled to increase the image resolution. Subsequently, the processed shallow features and deep features are fused, and features are extracted through a 3×3 convolution to achieve the segmentation of the sample image.

[0096] The Atrous Spatial Pyramid Pooling module is an important branch of the DeepLabv3+ semantic segmentation network. It uses atrous convolutions with different dilation rates to increase the receptive field of the network and achieve multi-scale feature extraction. The ASPP module consists of 4 parallel convolutional and pooling layers. Among them, the 4 parallel convolutions are a 1×1 standard convolution and 3 3×3 atrous convolutions with dilation rates of 6, 12, and 18 respectively. The ASPP module fuses the results obtained from the atrous convolutions to reduce the information loss rate and extract context semantic feature information in multiple proportions. The structure of the ASPP module is as Figure 5 shown.

[0097] Let H h n (x) represent the convolution operation, h represent the convolution kernel size, n represent the atrous size, and Q(x) represent the pooling. The output of ASPP is expressed as:

[0098] Y = H1 1 (x) + H3 6 (x) + H3 12 (x) + H3 18 (x) + Q(x)

[0099] In this embodiment, the excellent performance of DeepLabV3+ in image segmentation is utilized and further improved to enhance the image segmentation accuracy and data processing rate of the network model in potato root system segmentation. The improved network model is as shown in Figure 6 Figure. Specifically, the improved image segmentation network selects the DeepLabV3+ network model with an encoder-decoder structure as the basic model, and replaces the backbone network Xception in the basic model with the MobileNetV2 network. In the network model, the CARAFE upsampling module is used to replace the Upsample by 4 upsampling module in the basic network, and the high-level features output by the encoder and the non-output results of the decoder are upsampled. And a CBAM attention mechanism module is added respectively after the low-level features output by the DCNN and the feature connection layer of the encoder.

[0100] The CARAFE upsampling method is a general, lightweight, and efficient operator that can aggregate context information within a large receptive field, can perceive specific content, thereby dynamically generating an adaptive kernel, and has a fast calculation speed and is easy to integrate into modern network architectures. It mainly consists of an upsampling kernel prediction module and a feature recombination module. At each position, CARAFE can use the underlying content information to predict the recombination kernel and recombine the features within a predefined nearby area.

[0101] The upsampling kernel prediction module is responsible for generating the recombination kernel in a content-aware manner. Assuming the upsampling ratio is σ, for an input feature map with a size of C×H×W, a 1×1 convolutional layer is used to compress the channels, compressing it to C m , obtaining a C m ×H×W feature map, which can reduce the subsequent calculation amount; then, through a convolutional kernel with a size of k encoder ×k encoder to encode the above content to generate the recombination kernel, obtaining a feature map with a size of ; the above-obtained feature map is recombined into and normalized using the softmax function so that the sum of the weights of the convolutional kernel is 1. Among them, C is the length of the input feature map, H and W are the height and width of the input feature map respectively; C m is the number of input channels; is the number of output channels.

[0102] For each recombination kernel W L′ , the feature recombination module recombines the features within the local area through a weighted summation operator. For the position L′ and the square area N=(X L ,k up ) centered at L=(i,j), the following formula is used to calculate the recombined feature at L′:

[0103]

[0104] In the above formula, W L′ is the feature recombination kernel; r is k up 2; X is the original input feature map.

[0105] By performing a dot product operation between the input feature map and the predicted upsampling kernel, the upsampling result is obtained. The recombined feature map has stronger semantics than the original input feature map and can better focus on the relevant point information within the local area. Compared with traditional upsampling methods, CARAFE can use an adaptively optimized recombination kernel at different positions, with very small parameter and computational amounts, which helps to improve the performance of the upsampling operator.

[0106] Aiming at the problems of slow fitting speed of the standard Deeplabv3+ model for image segmentation and low accuracy of target edge segmentation, this embodiment introduces the CBAM attention mechanism as shown in Figure 7 to process the high-level and low-level feature layers, enabling the model to give different weights and attentions to different parts of the input image, and enhancing the sensitivity and accuracy of the semantic segmentation network for feature extraction.

[0107] The CBAM attention mechanism consists of a channel attention module and a spatial attention module. Among them, the channel attention module focuses on which features of the image are more meaningful, while the spatial attention module focuses on which regions of the features are more meaningful. This module not only pays attention to the proportion of each channel, but also to the proportion of each pixel point, and can be adaptively optimized according to the features of the input image. In addition, a major advantage of the CBAM module is that it is lightweight and can be seamlessly integrated into any neural network for plug-and-play.

[0108] The input feature layer of the CBAM module undergoes pooling operations through the channel attention module to obtain the weights of each channel of the input feature layer, and applies them to the spatial attention module. The spatial attention module takes the maximum and average values on the channels of each feature point, and then through the same operations as the channel attention, obtains the weights of each feature point of the input feature layer. Finally, the weights are multiplied with the original input feature layer and, after convolution processing, deep features containing multi-scale context information are obtained. In the decoder part, the extracted original features are fed into the CBAM module, and after the same processing, shallow features containing multi-scale context information are obtained. Finally, operations such as upsampling and convolution are performed to extract features from the fused image, realizing the accurate segmentation of potato root system images.

[0109] Meanwhile, to reduce the model parameter amount and improve the training speed, this embodiment uses the lightweight MobileNetV2 network as the backbone network of the model to replace the backbone network Xception in the original network model.

[0110] The MobileNetV2 network model introduces an inverted residual module and a linear bottleneck layer on the basis of using depthwise separable convolutions, greatly reducing the number of model parameters and thus making the network converge faster. This feature extraction network first obtains features of the same dimension through 3×3 depthwise convolution and ReLU6 activation function to prevent the nonlinear layer from destroying too much feature information. Then, it processes through 1×1 convolution and ReLU6 to obtain the downsampled features, and finally uses 1×1 convolution for upsampling. The inverted residual module is mainly used to improve the effective transmission of multi-layer feature information and enhance the network's feature extraction ability. For this module, the input first goes through 1×1 convolution for upsampling, then extracts features through 3×3 depthwise convolution, and finally uses 1×1 convolution to downsample the features to obtain feature information.

[0111] The working process of the improved image segmentation network in this embodiment includes the following steps: First, the input sample image extracts image features through the MobileNetV2 backbone network in the encoder part. Second, it enters the atrous spatial pyramid pooling module to obtain image spatial feature information and transmits it to the CBAM module. Then, the input feature layer of the CBAM module undergoes pooling operations through the channel attention module to obtain the weights of each channel of the input feature layer, and applies them to the spatial attention module. The spatial attention module takes the maximum and average values on the channels of each feature point, and then through the same operations as the channel attention, obtains the weights of each feature point of the input feature layer. Then, multiply this weight with the original input feature layer, and after convolution processing, obtain the deep features containing multi-scale context information. Next, in the decoder part, the extracted original features are sent into the CBAM module, and after the same processing, obtain the shallow features containing multi-scale context information. Finally, through operations such as CARAFE upsampling and convolution, feature extraction is performed on the fused image to accurately segment the potato root system part in the input sample image.

[0112] In the image segmentation module proposed in this embodiment, as Figure 8 shown, the training method of the image segmentation network is as follows:

[0113] S1: Select the true root system samples of the potato plants to be measured. After preprocessing the root system samples, lay them flat evenly on the sample placement board of the image acquisition device.

[0114] S2: Replace different root system samples respectively, and adjust the morphology, position of the root system samples, and the shooting parameters of the camera in the image acquisition device to obtain a large number of different root system sample images, constituting the required sample dataset.

[0115] S3: Use the dataset augmentation method to expand the number of sample images in the sample dataset, and then divide the sample dataset into a training set and a validation set.

[0116] S4: Annotate the characteristic parts of potato roots in the training set using the Labelme image annotation tool; the annotated training set includes 2 semantic classifications, namely the foreground category and the background category.

[0117] S5: Use the sample images in the pre-annotated training set to train the designed image segmentation network: The training process is as follows:

[0118] S51: Use the following mean intersection over union as the metric to evaluate the performance of the segmentation model, and the calculation formula is as follows:

[0119]

[0120] In the above formula, k is the total number of semantic categories; TP represents the accuracy of the model predicting root pixel values, TN represents the accuracy of the model predicting non-root pixel values; FP represents the false positive rate of the model; FN represents the false negative rate of the model predicted;

[0121] S52: Preset the training parameter settings in the training process as follows: Use a dynamic learning rate, set the initial learning rate value to 0.1, set the batch size to 16, and train for 60 epochs.

[0122] S53: Use the sample images in the training set with added annotations to train the image segmentation network; among them, before each round of training, randomly rearrange the order of the sample images in the dataset.

[0123] S54: End the training process of the network model after reaching the preset number of epochs, and verify the trained network model through the validation set; retain the model parameters of the trained image segmentation network.

[0124] In the training stage of the image segmentation network, the dataset augmentation methods used in step S3 include image rotation processing, image mirroring processing, and image enhancement processing; the methods of image enhancement processing include brightness change, sharpness change, and image blurring.

[0125] Figure 9 Is a case image for a series of mirroring and rotation processing of the original image. Figure 9 Part a in it is the original image, part b is the image obtained by horizontally mirroring the original image. Part c is the image obtained by vertically mirroring the original image data. D is the image obtained by rotating the original image counterclockwise by 90°, part e is the image obtained by rotating the original image counterclockwise by 180°, and part f is the image obtained by rotating the original image counterclockwise by 270°.

[0126] Figure 10 Is a case image for a series of image enhancement processing of the original image.Figure 10 In part a is the original image, part b is the high - brightness image obtained by enhancing the brightness of the original image. Part c is the low - brightness image obtained by reducing the brightness of the original image. Part d is the image obtained by enhancing the sharpness of the original image. Part e is the image obtained by reducing the sharpness of the original image. Part f is the image obtained by performing Gaussian blur processing on the original image.

[0127] In the coefficient query module of the image processing device, as Figure 11 shown, the method for generating the pixel conversion coefficient table is as follows:

[0128] (1) Obtain the image of a real calibrated ruler placed on the operation platform through the image acquisition device.

[0129] (2) Perform median filtering on the image of the ruler, and use the iterative segmentation method to segment the region of interest, find the centimeter marks on the ruler, and use morphological transformation to remove the millimeter marks.

[0130] (3) Obtain the horizontal projection of the centimeter marks on the ruler through Radon transform, and finally determine the intermediate interval value D between adjacent centimeter marks on this projection. i .

[0131] (4) According to the correspondence between the intermediate interval value D between adjacent centimeter marks i and the pixels in the image, calculate the conversion factor between "pixel - millimeter", that is, the pixel conversion coefficient B. The calculation formula is as follows:

[0132]

[0133] In the above formula, P d is the pixel value between adjacent centimeter marks; D i is the intermediate interval value between adjacent centimeter marks; the unit of the pixel conversion coefficient B is: pixel / mm.

[0134] (5) Adjust the shooting state of the image acquisition device, and repeat steps (1) - (4) to determine the pixel conversion coefficients corresponding to different shooting states, and record the mapping relationship between the shooting information and the pixel conversion coefficients to obtain the required pixel conversion coefficient table.

[0135] In this embodiment, Figure 2What is given is only the simplest scheme of the image acquisition system. In further improvements, the bracket in the image acquisition device is an electric lifting bracket; the fill light is a shadowless lamp system with multiple light sources. The fill light is located directly above the sample placement plate, and each light source is distributed in a ring shape. The camera in the image acquisition device is located at the center of the light sources. The illumination area of the fill light is the sample area in the lower sample stage, so that the brightness of different positions in the sample area is kept uniform when imaging in the camera. The light sources in the fill light are cold light lamps with a color temperature close to natural light.

[0136] In addition, the image acquisition device also includes a foldable light shield. In the unfolded state, the light shield is in a sleeve shape and covers the outer periphery of the sample area of the sample placement plate; the light shield is detachably connected to the operation platform, and a soft light film is provided on the inner wall of the light shield.

[0137] In the product scheme of the digital measurement system provided in this embodiment, the image processing device further includes a display screen and input devices such as a mouse and a keyboard. The image processing device also runs a human-computer interaction system. The digital measurement system interacts with the operator through the human-computer interaction system to enable a technician to manipulate the operation process of the digital measurement system.

[0138] Performance Test

[0139] In order to verify the effectiveness of the solution of the present invention and test the performance such as the measurement accuracy of the designed digital measurement system, the following performance test scheme is formulated in this embodiment.

[0140] In the performance test scheme, the technician built the corresponding digital measurement system, used the roots of large potatoes collected from Chayouzhongqi and Siziwangqi in Inner Mongolia as samples, collected sample images for the training of the network model, and used some of the collected samples as the measurement targets to test the performance of the digital measurement system.

[0141] 1. Construction of the digital measurement system

[0142] (1) Image acquisition device

[0143] The simple potato root image acquisition device built in this experiment mainly consists of a high-definition camera, a laptop computer, a measurement tablet, and a bracket for fixing the relative positions of each device. This image acquisition device is as Figure 12As shown in the figure, specifically, a camera with the model MV-CH050-10UC is fixed on the top of the acquisition platform. The maximum frame rate of the camera is 74 frames / s, and the resolution is 2448×2048 pixels. After each sampling, the root system is rinsed and placed on the measurement plate. The camera is fixed directly above the acquisition platform according to the preset camera position, and the root system image is collected. To ensure a good contrast between the root system image and the background, in this embodiment, a black cloth is covered on the measurement plate as the background to enhance the image contrast. No fill light is used in this image acquisition device, but sufficient indoor light is ensured during image acquisition. Considering the influence of indoor lighting, a light shield is used to block the light on the top of the platform to eliminate the influence of indoor lighting on the image acquisition process.

[0144] (2) Data processing device

[0145] In this experiment, the data processing device built uses the laptop in the image acquisition device as the hardware platform for data processing. And a computer system for measuring the length of potato root systems is established using the GUI interface of Matlab. This computer system contains a computer program corresponding to a complete data processing process with multiple different functional modules such as a data acquisition module, an image segmentation module, a coefficient query module, and a length calculation module. The human-computer interaction interface of the computer system established in this embodiment is as Figure 13 and Figure 14 shown. Figure 13 is the login interface for measuring the length of potato root systems, Figure 13 is the calculation interface for measuring the length of potato root systems.

[0146] 2. Production of sample images and data sets

[0147] (1) Acquisition of root system images

[0148] To collect the root system samples of potatoes, technicians randomly dug several soil blocks with an area of 180 cm×90 cm in the experimental field, and the depth of the soil blocks was 120 cm. Each soil block contained 6 potato plants with uniform growth. Among them, the 4 complete plants in the middle were selected, and sampling was carried out layer by layer from the soil surface downward at intervals of 10 cm. After the root systems in each soil layer were dug out together with the soil of that layer, they were passed through a 20-mesh sieve. The collected root systems were put into a mesh bag, impurities were removed using tweezers, rinsed with water, and then the surface moisture of the root system samples was blotted dry with filter paper for measuring the root system length.

[0149] In this experiment, by taking pictures of different root system samples at different angles and with different parameters, 976 different sample images were obtained. The above data was used as the original data set for training the image segmentation network in the data processing device.

[0150] (2) Amplification of the original data set

[0151] When training the image segmentation network for potato roots, root images are required as the model training set. To avoid overfitting of the model caused by insufficient training samples, image augmentation is usually used to increase the sample data volume. On this basis, an image labeling tool is used to label the characteristic parts of the roots to produce a data set for model training.

[0152] In this experiment, each image in the original data set was successively subjected to horizontal mirroring, vertical mirroring, 90° counterclockwise rotation, 180° counterclockwise rotation, 270° counterclockwise rotation, brightness enhancement, brightness reduction, high sharpness processing, low sharpness processing, and Gaussian blur processing. At the same time, the above method expanded the number of samples of the original data by 10 times. The number of images in the data set increased from the original 976 to 9760. In this experiment, the 9760 sample images were divided into: 5856 training set images (60%), 1952 validation set images (20%), and 1952 test set images (20%).

[0153] (3) Production of the training set

[0154] In this experiment, the Labelme image labeling tool was used to label the characteristic parts of potato roots so that the image segmentation network model could accurately learn the root characteristics of each category. According to the requirements of the semantic segmentation model, the data set includes 2 semantic classifications, namely 1 foreground category and 1 background category. Using the potato root structure framework as the learning feature for annotation, the json file data obtained from the annotation was converted into the VOC2007 format. This format of the data set includes three parts: ImageSets, JPEGImages, and SegmentationClass.

[0155] Under Segmentation in the ImageSets file, there are three sub-files, namely the train.txt file representing the training set, the val.txt file representing the validation set, and the trainval.txt file summarizing the training and validation sets. The JPEGImages folder stores all the original RGB root images, stored in the jpg format, and the images are named with a 6-digit number. The SegmentationClass folder stores the label files, that is, the potato root label images processed by the Labelme image labeling tool. The Labelme image labeling tool was used to label the characteristic parts of potato roots. As Figure 15 shown in the figure, the left side is the image annotation interface. In the middle image, the light-colored part represents the root label part, and the remaining dark-colored part is the background label, that is, 2 semantic classifications. The image on the right is the label image represented in grayscale.

[0156] 3. Training of the image segmentation network

[0157] (1) Model evaluation metrics

[0158] The mean intersection over union (mIoU) is used to evaluate the performance of the segmentation model. mIoU represents the ratio of the intersection to the union of the true pixel values and the predicted pixel values for each class. The calculation formula is as follows:

[0159]

[0160] In the above formula, k is the total number of semantic classes; TP represents the accuracy of the model predicting the root pixel value, that is, predicting as roots and actually being roots; TN represents the accuracy of the model predicting not to be the root pixel value, that is, predicting not to be roots and actually not being roots; FP represents the model prediction error, that is, predicting as roots but actually not being roots, which can also be expressed as the false positive rate; FN represents the model prediction error, that is, predicting not to be roots but actually being roots, which can also be expressed as the false negative rate.

[0161] (2) Model training environment configuration

[0162] In this experiment, a deep learning workstation with a 24-core Intel Xeon Platinum 8168 processor, 128G of memory, a main frequency of 2.7GHz, an NVIDIA Quadro P6000 graphics card, and 24GB of video memory is selected to train the segmentation model. The pytorch1.2.0 framework is used to build and adjust the parameters of the segmentation model in this paper. The model training environment configuration is shown in Table 1:

[0163] Table 1: Model training environment configuration table

[0164]

[0165]

[0166] The improved image segmentation network in this experiment is trained, and the model parameters are adjusted according to the training results. In model training, the selection of the learning rate is extremely important. A larger learning rate causes the weight parameters to cross the optimal value, while a smaller learning rate slows down the convergence of the training. To select a more appropriate learning rate, different values of the learning rate are often experimented with during the training process, which is computationally intensive and time-consuming. Therefore, this experiment uses a method of dynamically adjusting the learning rate over time to train the network.

[0167] Set the initial learning rate value to 0.1, set the batch size to 16, and train for 60 epochs. To ensure that the data seen in the same batch is different in different rounds of the model, the dataset is randomly shuffled before each training. This can not only improve the convergence speed of the model but also improve the prediction results of the model on the test set. After the model training is completed, open the model training log, which contains the change values of the loss functions of the training set and the validation set.

[0168] (3) Evaluation of the training results of the network model

[0169] In this experiment, the loss value curve of the image segmentation network in the training stage is as Figure 16 shown. Analyzing Figure 16 the data in it shows that: after the network model starts training, the loss value decreases gradually with the increase of epochs. The loss function values of the training set and the validation set drop to 0.103 and 0.111 respectively at the 50th epoch; then they gradually tend to be stable, and the loss function values finally converge to 0.097 and 0.103 at the 60th epoch. Therefore, with the increase of the training times, the loss values of the training set and the validation set both tend to be around 0.1, indicating that the model has good convergence of the loss function, which will be beneficial to improving the root image segmentation accuracy.

[0170] 4. Comparison of the image segmentation model adopted in this example with other solutions

[0171] (1) Comparison of backbone networks

[0172] To verify the effectiveness of the feature extraction network in the improved image segmentation network based on DeepLabV3+ designed in this experiment, control experiments are set respectively while keeping other parameters the same. This example is compared with the model constructed based on the ResNet50 feature extraction backbone network, and the results are shown in Table 2.

[0173] Table 2: Comparison of the model performances of different feature extraction backbone networks

[0174]

[0175] Analysis of the data in the figure shows that: Different feature extraction backbone networks can effectively segment potato root images. However, the training time of MobileNetV2 is only 10.2h, which is 3.3h less than that of ResNet50. In terms of segmentation performance, the MIoU of MobileNetV2 used in this example can reach 92.17%, which is 2.79 percentage points higher than that of ResNet50. The results in terms of model segmentation effect and training efficiency show that the feature extraction network based on MobileNetV2 has the best effect and can be used as the backbone network of the potato root image segmentation model in this experiment.

[0176] (2) Comparison of target segmentation effects

[0177] To verify the performance of the semantic segmentation model improved based on DeepLabv3+ in this example, this experiment specifically designed a control experiment. The root image segmentation process was carried out using the sample images of simple main roots and the sample images containing complex lateral roots respectively, and the segmentation results of the network model in this example were compared with the standard DeeplabV3+ segmentation model. The results are as Figure 17 and Figure 18 shown.

[0178] In Figure 17 the simple root samples, although the standard DeepLabv3+ semantic segmentation model can accurately segment the potato roots from the image, compared with this example, the standard DeepLabv3+ semantic segmentation model still has missegmentation in some parts of the image. The yellow boxes in the figure indicate some missing root pixels, while the blue boxes indicate the soil particles accompanying during the experiment.

[0179] In Figure 18 the complex root samples, the original image contains more lateral roots and root hairs, and the pixel points near the lateral roots and root hairs are sparser. Comparing the segmentation results in the yellow boxes in the figure, it is found that the standard DeepLabv3+ semantic segmentation model loses some root pixels and detailed features, while the improved network model in this example solves these problems well. The improved DeepLabv3+ has better comprehensive analysis ability for root images without training.

[0180] In addition, on the premise that other conditions remain the same, this experiment also calculated the MIOU values of the model in this example and the standard DeepLabv3+ model on the root images in the validation set. The statistical results are shown in Table 3.

[0181] Table 3: Performance comparison of different segmentation methods on the validation set

[0182]

[0183] Analysis of the data in the above table shows that: The segmentation effect of the improved network model provided in this experiment is the best, with the MIoU reaching 93.52%, which is 2.77 percentage points higher than that of the standard DeepLabv3+ model. The improved DeepLabv3+ model provided in this experiment can be used as the image segmentation network model for potato roots.

[0184] 5. Refinement of Image Segmentation Results

[0185] In the digital measurement system for potato root length provided in this embodiment, after the length calculation model obtains the segmented local root image, the Hilditch refinement algorithm is first used to optimize the pixels of the target image, that is, to further refine the identified root part to ensure that the identified root part is more detailed and avoid the interference of background pixels. In this experiment, the pixel optimization result of the Hilditch refinement algorithm for the Figure 17 segmented local root image in Figure 19 is shown as follows. Analysis Figure 19 of the results shows that: Almost all the valid information of the root length is retained in the refined root image, which provides a guarantee for the calculation of the root length.

[0186] 6. Comparison between Digital Measurement Scheme and Manual Scheme

[0187] To verify the measurement accuracy of the digital scheme in this experiment, 50 images were randomly selected from the root dataset during the performance measurement, and the total root length was calculated and compared with the results of manually measured root lengths. The fitting curve shown in Figure 20 was generated according to the measurement results of different schemes. Analysis Figure 20 of the data in it shows that: The pearson correlation coefficient of the curve fitting corresponding to the measurement results of the two different measurement schemes reaches 0.967, which indicates that the digital measurement scheme proposed in this example has a high credibility, and its measurement accuracy can fully compare with the manual measurement scheme and can be applied in practice.

Claims

1. A digital measurement system for the root length of potatoes, characterized in that, It includes: An image acquisition device, which includes a sample stage, a bracket, a camera, and a fill light; the sample stage includes a horizontal operation platform; the center of the operation platform contains a detachable transparent sample placement plate, and the upper surface of the sample placement plate is a sample area for placing potato root samples to be measured; the inside of the sample placement plate contains a sandwich space for placing a solid-color background plate; the camera and the fill light are fixedly connected above the sample stage through the bracket; the camera is used to collect sample images of the measurement target on the background plate along the orthographic projection direction, and the fill light is used to provide uniform ambient light during the camera shooting process; and A data processing device, which is communicatively connected to the image acquisition device and is used to manipulate the image acquisition device; the data processing device includes a data acquisition module, an image segmentation module, a coefficient query module, and a length calculation module; the data acquisition module is used to obtain a sample image captured by a camera and extract the shooting information corresponding to the sample image when it is captured; the image segmentation module contains an image segmentation network for identifying a root system sample and segmenting the identified root system sample; the image segmentation network is trained based on an improved DeepLabV3+ network model; the image segmentation module is used to perform feature recognition and image segmentation on the input sample image, and then output a target image that only contains the root system part after segmentation; the coefficient query module is preset with a pixel conversion coefficient table for characterizing the mapping relationship between pixels and the true scale in the sample image determined by the camera under different shooting states; and is used to query the corresponding pixel conversion coefficient according to the shooting information of the sample image extracted by the data acquisition module B ; the length calculation module is used to first optimize the pixels of the target image through the Hilditch thinning algorithm and calculate the root system length represented by the pixels corresponding to the potato root system in the sample image P L ; then combine the pixel conversion coefficient obtained by the coefficient query model B , and use the formula R L = B · P L to calculate the true root system length R L ; The image segmentation network selects the DeepLabV3+ network model with an encoder-decoder structure as the basic model, and replaces the backbone network Xception in the basic model with the MobileNetV2 network; in the network model, the CARAFE upsampling module is used to replace the Upsample by 4 upsampling module in the basic network, and the high-level features output by the encoder and the non-output results of the decoder are upsampled; and a CBAM attention mechanism module is added after the low-level features output by the DCNN and the feature connection layer of the encoder respectively.

2. The digital measurement system for the root length of potatoes according to claim 1, characterized in that: The working process of the image segmentation network is as follows: First, the input sample image extracts image features through the MobileNetV2 backbone network in the encoder part; Second, enter the atrous spatial pyramid pooling module to obtain image spatial feature information and transmit it to the CBAM module; Then, the input feature layer of the CBAM module performs pooling operations through the channel attention module to obtain the weights of each channel of the input feature layer, and apply them to the spatial attention module; The spatial attention module takes the maximum value and the average value on the channels of each feature point, and then through the same operation as the channel attention, obtains the weights of each feature point of the input feature layer; then multiplies the weights by the original input feature layer and obtains deep features containing multi-scale context information after convolution processing; Next, in the decoder part, the extracted original features are sent into the CBAM module, and shallow features containing multi-scale context information are obtained after the same processing; Finally, feature extraction is performed on the fused image through CARAFE upsampling and convolution operations to accurately segment the potato root part in the input sample image.

3. The digital measurement system for the root length of potatoes according to claim 1, characterized in that: In the image segmentation module, the training method of the image segmentation network is as follows: S1: Select the true root samples of the potato plants to be measured, preprocess the root samples and evenly spread them on the sample placement plate of the image acquisition device; S2: Replace different root samples respectively, and adjust the morphology, position of the root samples, and the shooting parameters of the camera in the image acquisition device to obtain a large number of different root sample images to form the required sample data set; S3: Use the data set augmentation method to expand the number of sample images in the sample data set, and then divide the sample data set into a training set and a validation set; S4: Labelme image annotation tool is used to annotate the characteristic parts of potato roots in the training set; the annotated training set includes two semantic classifications, namely foreground class and background class; S5: The designed image segmentation network is trained with the sample images in the pre-annotated training set: the training process is as follows: S51: The following mean intersection over union is used as an index to evaluate the performance of the segmentation model, and the calculation formula is as follows: In the above formula, k is the total number of semantic categories; TP represents the accuracy rate of the model predicting root system pixel values, TN represents the accuracy rate of the model predicting non-root system pixel values; FP represents the false alarm rate of the model; FN represents the missed alarm rate predicted by the model; S52: The training parameters in the preset training process are set as follows: using a dynamic learning rate, setting the initial learning rate value to 0.1, setting the batch size to 16, and training for 60 epochs; S53: The sample images in the training set after adding annotations are used to train the image segmentation network; among them, the order of the sample images in the dataset is randomly rearranged before each round of training; S54: After reaching the preset number of epochs, the training process of the network model is ended, and the trained network model is verified through the validation set; the model parameters of the image segmentation network after training are retained.

4. The digital measurement system for potato root length according to claim 3, wherein: In step S3, the dataset augmentation methods adopted include image rotation processing, image mirror processing, and image enhancement processing; the ways of image enhancement processing include brightness change, sharpness change, and image blurring.

5. The digital measurement system for the root length of potatoes according to claim 1, characterized in that: In the coefficient query module, the generation method of the pixel conversion coefficient table is as follows: (1) An image of a real calibrated ruler placed on the operation platform is obtained through an image acquisition device; (2) Median filtering is performed on the image of the ruler, and an iterative segmentation method is used to segment the region of interest, find the centimeter marks on the ruler, and use morphological transformation to remove the millimeter marks; (3) Obtain the horizontal projection of the centimeter marks on the ruler through the Radon transform, and finally determine the intermediate interval value between adjacent centimeter marks on this projection D i ; (4)According to the intermediate interval value between adjacent centimeter marks D i and the corresponding relationship between pixels in the image, calculate the conversion factor between "pixels - millimeters", that is, the pixel conversion coefficient B , and the calculation formula is as follows: In the above formula, P d is the pixel value between adjacent centimeter marks; D i is the intermediate interval value between adjacent centimeter marks; the pixel conversion coefficient B is in the unit of: pixel / mm; (5) Adjust the shooting state of the image acquisition device, and repeat steps (1) to (4) to determine the pixel conversion coefficients corresponding to different shooting states, and record the mapping relationship between the shooting information and the pixel conversion coefficients to obtain the required pixel conversion coefficient table.

6. The digital measurement system for the root length of potatoes according to claim 1, characterized in that: The bracket in the image acquisition device adopts an electric lifting bracket; the fill light adopts a multi-light source shadowless lamp system; the fill light is located directly above the sample placement plate, and each light source is distributed in a ring shape, and the camera in the image acquisition device is located in the center of the light source; the illumination area of the fill light is the sample area in the lower sample stage, so that the brightness of different positions in the sample area is kept uniform when imaging in the camera; the light source in the fill light adopts a cold light lamp with a color temperature close to natural light.

7. The digital measurement system for the root length of potatoes according to claim 1, characterized in that: The image acquisition device also includes a foldable light shield, which is in a sleeve shape in the unfolded state and covers the outer periphery of the sample area of the sample placement plate; the light shield is detachably connected to the operation platform, and a soft light film is provided on the inner wall of the light shield.

8. The digital measurement system for the root length of potatoes according to claim 1, characterized in that: The image processing device includes a display screen and an input device; the image processing device also runs a human-computer interaction system, and the digital measurement system interacts with the operator through the human-computer interaction system to realize the operation process of the digital measurement system by technicians.

9. A digital measurement method for the root length of potatoes, characterized in that: The digital measurement system of the potato root system described in any one of claims 1-8 is a product designed using the concept of the digital measurement method for the length of the potato root system; the digital measurement method for the length of the potato root system includes the following steps: S01A: Set up an image acquisition device for obtaining a sample image of the target to be measured in the orthographic projection direction; S02A: Design an image segmentation network based on deep learning technology for extracting the target object to be measured from the acquired sample image; S03A: Take test images of a real ruler with different shooting parameters through the image acquisition device, and calculate the pixel conversion coefficient of the image acquisition device in different shooting states according to the object-image ratio in the test images; The pixel conversion coefficient is the conversion factor between the pixel representation length and the real length in the sample images taken by the image acquisition device under different shooting parameter conditions; S04A: After segmenting and flattening the potato root system sample to be measured, place it on the image acquisition device, adjust the shooting parameters of the image acquisition device to obtain a sample image with the best image quality, and determine the pixel conversion coefficient corresponding to the current shooting parameters; S05A: Input the sample image into the image segmentation network, and the image segmentation network performs feature extraction and image segmentation on the sample image to obtain the segmented root system local image; S06A: Use the Hilditch thinning algorithm to thin the obtained root system local image, remove the isolated interfering pixels in the root system local image, and output the pixel values corresponding to the root system local image; S07A: Obtain the root length corresponding to the pixel representation of all root hairs in the optimized local root image, and calculate the true root length through the following formula P L , and calculate the true root length through the following formula R L : R L = B · P L In the above formula, B is the pixel conversion coefficient between the pixel length and the true length of the target in the known sample image.

Images