Rice combine harvester grain loss detection method based on SSD-FFNet

Abstract

The invention relates to the field of agricultural grain loss detection, in particular to an SSD-FFNet-based grain loss detection method for a rice combine harvester. Comprising the following steps: sampling at a discharge port of the combine harvester; screening is conducted, and long straw is screened away; small-size impurities are blown off as much as possible by using an air blowing device, and grain images are collected; detecting rice grains by using a target detection algorithm SSD-FFNet; then, the grain loss rate is calculated according to the detected number of complete and incomplete grains; afterwards, the steps are repeatedly executed, and multiple sets of grain loss rates are obtained; and carrying out average calculation on multiple groups of data to obtain a final grain loss rate. According to the scheme, accurate detection and real-time monitoring of the grain loss rate are achieved, immediate feedback is provided for operators of the rice combine harvester, the operators are helped to adjust operation parameters of the harvester, grain waste is reduced, the harvesting quality and economic benefits are improved, important reference is provided for intelligent upgrading of the harvester, and the method is worthy of popularization and application. And the progress of the modern agricultural equipment technology is promoted.

Description

Technical Field

[0001] The present invention relates to the field of agricultural grain loss detection, and in particular to a rice combine harvester grain loss detection method based on SSD-FFNet. Background Art

[0002] As an important food crop, the yield and quality of rice are directly related to food security and farmers' economic benefits. During the rice harvesting process, grain loss detection is one of the key links in modern agricultural harvesting operations, which is related to harvesting efficiency, economic benefits and full utilization of food resources. However, in the actual harvesting process, due to factors such as mechanical operation, weather conditions or crop characteristics, rice grains are prone to damage, missing or mixing. The presence of these incomplete grains significantly reduces the overall quality of rice, and in severe cases may also affect subsequent processing and market value. Therefore, how to efficiently and accurately detect the loss of rice grains and take timely measures to reduce losses is of great significance to improving agricultural production efficiency and ensuring food quality.

[0003] At present, rice grain loss detection mainly includes two methods: manual detection and technology-based automatic detection. The manual detection method has problems such as low efficiency, poor accuracy, and inability to achieve online real-time detection. In addition, the detection results are easily affected by subjective factors such as fatigue and lack of experience of the detection personnel, which makes it difficult to meet the needs of modern agriculture for precision and intelligence. Technology-based automatic detection mainly relies on traditional image recognition technology and piezoelectric methods. Although traditional image recognition technology has improved the detection efficiency to a certain extent, it often requires pre-setting feature extraction algorithms, which makes it difficult to flexibly respond to the diversity of rice grains of different varieties and different growth conditions. Especially in the environment of high-speed operation of combine harvesters, traditional methods often fail to meet accurate detection needs. The piezoelectric method uses piezoelectric sensors to sense the vibration generated by the falling of grains and convert it into electrical signals to detect the loss, but it has the disadvantages of being easily disturbed by external vibrations, limited detection range, greatly affected by grain characteristics, complex signal processing, high equipment maintenance costs, and difficult installation and debugging.

[0004] With the rapid development of deep learning technology, especially the application of convolutional neural networks in the field of target detection, new ideas have been provided for rice grain loss detection. Deep learning can automatically learn and extract complex features in images without the need for manual design of rules, thereby improving the accuracy and adaptability of detection. However, deep learning models also face challenges when dealing with rice grain loss detection. In the scenario of high-speed operation of combine harvesters, how to quickly and accurately detect incomplete rice grains has become a technical problem that needs to be solved urgently. Therefore, in order to design a deep learning model suitable for rice grain loss detection, the following conditions must be met: (1) high real-time performance, that is, the model can quickly process image data during the operation of the combine harvester, realize instant feedback, and ensure that the detection does not affect the harvesting efficiency; (2) strong small target detection ability, the model should be able to effectively identify and locate rice grains with small size and unclear features in the image, and improve the accuracy and comprehensiveness of detection; (3) strong robustness, the model can maintain stable detection performance under different lighting conditions, crop varieties and complex harvesting environments. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a rice combine harvester grain loss detection method based on SSD-FFNet. The detection technology is a rice combine harvester grain loss detection method based on an improved single shot multibox detection (SSD) target detection algorithm SSD-FFNet (Single Shot MultiBox Detector with FeatureFusion Net) feature fusion network. First, sampling is performed at the discharge port of the combine harvester; then, a sieve is used to screen out the longer straw and retain the grains and small-volume debris; then, a blowing device is used to blow away the small-volume debris as much as possible; then, an industrial camera is used to collect the image of the grain after screening and shaving; then, the target detection algorithm SSD-FFNet is used to detect the number of complete grains and the number of incomplete grains in the image; then, the grain loss rate is calculated according to the number of complete and incomplete grains detected; thereafter, the above steps are repeated to obtain multiple groups of grain loss rates; finally, the multiple groups of data are averaged to obtain the final grain loss rate. The rice combine harvester grain loss detection method includes sampling, screening, blowing, image acquisition and image detection technology, which realizes accurate detection and real-time monitoring of the grain loss rate and provides a reference for the operation status of the rice combine harvester.

[0006] Specifically, the present invention provides a rice combine harvester grain loss detection method based on SSD-FFNet, which is characterized by comprising the following steps:

[0007] Real-time sampling of grains at the discharge port of the combine harvester and pre-treatment;

[0008] Collect images of kernels after screening and peeling;

[0009] The target detection algorithm SSD-FFNet is used to detect the number of complete and incomplete rice grains in the image; the grain loss rate is calculated based on the number of grain losses;

[0010] Repeat the above steps to obtain multiple groups of grain loss rates; average the multiple groups of data to obtain the final grain loss rate.

[0011] Furthermore, the sampled grains are screened with a sieve to remove the long straw and retain the grains and small-volume debris; and then the small-volume debris is blown away with a blowing device.

[0012] Furthermore, sampling is performed using a roulette wheel periodic sampling with a sampling period of w minutes, and multiple samplings are performed at different time periods to avoid data deviations caused by local or instantaneous losses.

[0013] Furthermore, the size of the collected grain image I is normalized to obtain a normalized image I 1 ;

[0014] For the normalized image I 1 Divide into 6 feature prediction layers for feature extraction, and obtain 6 feature maps;

[0015] Apply prediction boxes of different sizes to each position on the six feature maps; perform convolution on each prediction box to obtain the category score and coordinate offset of each position;

[0016] The number of complete and incomplete kernels is obtained based on the category score and coordinate offset.

[0017] Furthermore, the grain image I is normalized. Specifically, the grain image I is a pixel matrix, and the number of rows of the pixel matrix is ​​the height h of the grain image I. 1 , the number of columns of the pixel matrix is ​​the width w of the grain image I 1 , use upsampling or downsampling methods to normalize the size of the grain image I to an image I of 300×300×3 1。

[0018] Furthermore, for the normalized image I 1 Perform feature extraction, including,

[0019] 1) The first part of the feature extraction module:

[0020] The first feature prediction layer is created through convolution and pooling to extract features. This part consists of thirteen layers, including two convolutional layers of size 3×3 and 64 convolution kernels, two convolutional layers of size 3×3 and 128 convolution kernels, three convolutional layers of size 3×3 and 256 convolution kernels, three convolutional layers of size 3×3 and 512 convolution kernels, and three maximum pooling layers. The input matrix size of this part is 300×300×3 image I 1 After convolution and pooling operations, the final output size is 38×38×512 feature map y 1 ;

[0021] 2) The second part: feature extraction module

[0022] The second feature prediction layer is created through convolution and pooling to extract features. This part consists of a seven-layer network, including three convolutional layers of size 3×3 and 512 convolution kernels, one convolutional layer of size 3×3 and 1024 convolution kernels, one convolutional layer of size 1×1 and 1024 convolution kernels, and two maximum pooling layers. The input matrix size of this part is the feature map y of 38×38×512. 1 After convolution and pooling operations, the final output feature map y is 19×19×1024 in size. 2 ;

[0023] 3) The third part: feature extraction module

[0024] The third feature prediction layer is created through convolution and pooling to extract features. This part consists of a two-layer network, including a convolution layer with a size of 1×1 and 256 convolution kernels, and a convolution layer with a size of 3×3 and 512 convolution kernels. The input matrix size of this part is the feature map y of 19×19×1024. 2 After convolution and pooling operations, the final output feature map y is 10×10×512 in size. 3 ;

[0025] 4) The fourth part: feature extraction module

[0026] The fourth feature prediction layer is created through convolution and pooling to extract features. This part consists of two branches, which are composed by splicing by adding the number of channels. Branch 1 consists of a two-layer network, including a convolution layer with a size of 1×1 and 128 convolution kernels, and a convolution layer with a size of 3×3 and 128 convolution kernels. Branch 2 consists of an adaptive pooling layer and a convolution layer with a size of 1×1 and 128 convolution kernels. The input matrix size of branch 1 is a feature map y of 10×10×512 3 , the input matrix size of branch 2 is the feature map y of 38×38×5121 After convolution, pooling and concatenation, the final output is a feature map y with a size of 5×5×256 4 ;

[0027] 5) The fifth part: feature extraction module

[0028] The fifth feature prediction layer is created through convolution and pooling to extract features. This part consists of two branches, which are composed by splicing by adding the number of channels. Branch 1 consists of a two-layer network, including a convolution layer with a size of 1×1 and 128 convolution kernels, and a convolution layer with a size of 3×3 and 128 convolution kernels. Branch 2 consists of an adaptive pooling layer and a convolution layer with a size of 1×1 and 128 convolution kernels. The input matrix size of branch 1 is the feature map y of 5×5×256 4 , the input size of branch 2 is the feature map y of 19×19×1024 2 After convolution, pooling and concatenation, the final output is a feature map y with a size of 3×3×256 5 ;

[0029] 6) Part 6 Feature Extraction Module

[0030] The sixth feature prediction layer is created through convolution and pooling to extract features. This part consists of two branches, which are composed by splicing by adding the number of channels. Branch 1 consists of a two-layer network, including a convolution layer with a size of 1×1 and 128 convolution kernels and a convolution layer with a size of 3×3 and 128 convolution kernels. Branch 2 consists of an adaptive pooling layer and a convolution layer with a size of 1×1 and 128 convolution kernels. The input matrix size of branch 1 is the feature map y of 3×3×256. 5 , the input size of branch 2 is a feature map y of 10×10×512 3 After convolution and pooling operations, the final output feature map y with a size of 1×1×256 is 6 .

[0031] Furthermore, prediction boxes of different sizes are applied to each position on the six feature maps. Specifically, in the generated feature map y 1 ,y 5 ,y 6 Construct three prediction boxes with different aspect ratios on each pixel in the prediction box. The aspect ratios of the prediction box include three ratios [1, 2, 1 / 2]; in the generated feature map y 2 ,y 3 ,y 4 Construct five prediction boxes of different scales and aspect ratios on each pixel in the feature map y. The aspect ratios of the prediction boxes include five ratios: [1, 2, 1 / 2, 3, 1 / 3]. 1-y 6 The scales of the prediction boxes generated above are [30, 60, 111, 162, 213, 264].

[0032] Furthermore, each prediction box is convolved to obtain the category score and coordinate offset of each position; specifically, the generated prediction boxes are convolved with two convolutional layers of size 3×3, one convolution outputs the category, i.e., the probability scores of complete and incomplete kernels, and the other convolution outputs the position coordinates (x, y, w, h) for regression.

[0033] Furthermore, the category and prediction box of each grain target are obtained according to the category score and coordinate offset; specifically,

[0034] The obtained grain category prediction frames are gathered together, and the non-maximum suppression (NMS) method is used to suppress a part of the overlapping or incorrect prediction frames. Position regression is performed based on the position coordinates generated above to generate the final prediction frame set, i.e., the grain detection result. The number is determined according to the number of detection frames, where the number of prediction frames for the category of incomplete grains is n, the number of prediction frames for the category of complete grains is m, and the number of prediction frames for all categories is N.

[0035] Furthermore, the mathematical model of the grain loss rate S is the ratio of the total number of lost grains to the total number of grains. The grain loss rate is calculated as shown in formula (1).

[0036] S=n / N

[0037] N=m+n (1)

[0038] Where n is the number of incomplete kernels; m is the number of complete kernels; and N is the number of all kernels.

[0039] This solution achieves accurate detection and real-time monitoring of grain loss rate, provides instant feedback to rice combine harvester operators, helps them adjust harvester operating parameters, reduces food waste, improves harvest quality and economic benefits, and provides important reference for intelligent upgrades of harvesters, promoting the advancement of modern agricultural equipment technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A flow chart of a rice combine harvester grain loss detection method based on SSD-FFNet provided by the present invention;

[0041] Figure 2 This is the structural diagram of the feature extraction part. DETAILED DESCRIPTION

[0042] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0043] refer to Figure 1 The present invention provides a rice combine harvester grain loss detection method based on SSD-FFNet, comprising the following steps:

[0044] Step 1: Sampling at the discharge port: The discharge port of the combine harvester is periodically sampled using a wheel, with a sampling period of w minutes, where w can range from 1 to 3;

[0045] Step 2: Screening: Use a sieve to screen out the longer straws and keep the grains and small debris;

[0046] Step 3: Blow away with air: Use the air blowing device to blow away small volume debris as much as possible;

[0047] Step 4: Image acquisition: Use an industrial camera to align the sampling platform to acquire the grain image to obtain image I;

[0048] Step 5: Grain detection. Specifically, the target detection algorithm SSD-FFNet is used to detect rice grains in the image (including the number of complete grains and incomplete grains), including the following steps:

[0049] Step 5.1: Perform normalization preprocessing on the size of the grain image I obtained in step 4 to obtain image I 1 ;

[0050] The grain image I is a pixel matrix, where the number of rows corresponds to the image height h 1 , the number of columns corresponds to the width w 1 . Due to the fixed input size of the CNN network, the original image needs to be standardized. Before feature extraction, the image is uniformly adjusted to a standard size of 300×300×3. For images of insufficient size, upsampling is used to increase the size of the original image. For images of oversized size, downsampling is used to reduce the size of the original image. Finally, an image I with a size of 300×300×3 is obtained. 1 .

[0051] Step 5.2: For image I 1 Perform feature extraction, which is divided into 6 feature prediction layers to obtain 6 feature maps;

[0052] A deep network is created through multi-layer convolution and pooling to extract features at different levels. The feature extraction module is divided into 6 feature prediction layers, resulting in 6 feature maps. Figure 2In the figure, Conv represents the convolution layer, 3×3 represents the size of the convolution kernel, 64 represents the number of convolution kernels, and 2 represents the number of convolution layers. Max Pooling represents the maximum pooling layer, 2×2 represents the size of the pooling filter, and Adaptive Pooling represents adaptive pooling.

[0053] Step 5.2.1: The first part of the feature extraction module

[0054] The first feature prediction layer is created through convolution and pooling to extract features. This part consists of thirteen layers, including two convolutional layers of size 3×3 and 64 convolution kernels, two convolutional layers of size 3×3 and 128 convolution kernels, three convolutional layers of size 3×3 and 256 convolution kernels, three convolutional layers of size 3×3 and 512 convolution kernels, and three maximum pooling layers. The input matrix size of this part is 300×300×3 image I 1 After convolution and pooling operations, the final output size is 38×38×512 feature map y 1 .

[0055] Step 5.2.2: The second part of the feature extraction module

[0056] The second feature prediction layer is created through convolution and pooling to extract features. This part consists of a seven-layer network, including three convolutional layers of size 3×3 and 512 convolution kernels, one convolutional layer of size 3×3 and 1024 convolution kernels, one convolutional layer of size 1×1 and 1024 convolution kernels, and two maximum pooling layers. The input matrix size of this part is the feature map y of 38×38×512. 1 After convolution and pooling operations, the final output feature map y is 19×19×1024 in size. 2 .

[0057] Step 5.2.3: The third part of the feature extraction module

[0058] The third feature prediction layer is created through convolution and pooling to extract features. This part consists of a two-layer network, including a convolution layer with a size of 1×1 and 256 convolution kernels, and a convolution layer with a size of 3×3 and 512 convolution kernels. The input matrix size of this part is the feature map y of 19×19×1024. 2 After convolution and pooling operations, the final output feature map y is 10×10×512 in size. 3 .

[0059] Step 5.2.4: The fourth part of the feature extraction module

[0060] The fourth feature prediction layer is created through convolution and pooling to extract features. This part consists of two branches, which are composed by splicing by adding the number of channels. Branch 1 consists of a two-layer network, including a convolution layer with a size of 1×1 and 128 convolution kernels, and a convolution layer with a size of 3×3 and 128 convolution kernels. Branch 2 consists of an adaptive pooling layer and a convolution layer with a size of 1×1 and 128 convolution kernels. The input matrix size of branch 1 is a feature map y of 10×10×512 3 , the input matrix size of branch 2 is the feature map y of 38×38×512 1 After convolution, pooling and concatenation, the final output is a feature map y with a size of 5×5×256 4 .

[0061] Step 5.2.5: The fifth part of the feature extraction module

[0062] The fifth feature prediction layer is created through convolution and pooling to extract features. This part consists of two branches, which are composed by splicing by adding the number of channels. Branch 1 consists of a two-layer network, including a convolution layer with a size of 1×1 and 128 convolution kernels, and a convolution layer with a size of 3×3 and 128 convolution kernels. Branch 2 consists of an adaptive pooling layer and a convolution layer with a size of 1×1 and 128 convolution kernels. The input matrix size of branch 1 is the feature map y of 5×5×256 4 , the input size of branch 2 is the feature map y of 19×19×1024 2 After convolution, pooling and concatenation, the final output is a feature map y with a size of 3×3×256 5 .

[0063] Step 5.2.6: Part 6 Feature Extraction Module

[0064] The sixth feature prediction layer is created through convolution and pooling to extract features. This part consists of two branches, which are composed by splicing by adding the number of channels. Branch 1 consists of a two-layer network, including a convolution layer with a size of 1×1 and 128 convolution kernels and a convolution layer with a size of 3×3 and 128 convolution kernels. Branch 2 consists of an adaptive pooling layer and a convolution layer with a size of 1×1 and 128 convolution kernels. The input matrix size of branch 1 is the feature map y of 3×3×256. 5 , the input size of branch 2 is a feature map y of 10×10×512 3 After convolution and pooling operations, the final output feature map y with a size of 1×1×256 is 6 ;

[0065] Step 5.3: Apply prediction boxes of different sizes to each position on the 6 feature maps. Specifically,

[0066] The feature map y generated in step 5.2 1 ,y 5 ,y 6 Construct three prediction boxes with different aspect ratios on each pixel in , and the aspect ratios of the prediction boxes include three ratios [1, 2, 1 / 2]. In the feature map y generated in step 5.2 2 ,y 3 ,y 4 Construct 5 prediction boxes of different scales and aspect ratios on each pixel in the feature map y. The aspect ratio of the prediction box includes five ratios: [1, 2, 1 / 2, 3, 1 / 3]. 1 -y 6 The scales of the prediction boxes generated above are [30, 60, 111, 162, 213, 264].

[0067] Step 5.4: Category prediction and position regression

[0068] The prediction boxes generated in step 5.3 are convolved with two convolutional layers of size 3×3. One convolution outputs the category, i.e., the probability scores of complete and incomplete kernels, and the other convolution outputs the position coordinates (x, y, w, h) for regression.

[0069] Step 5.5 Prediction box screening and correction

[0070] Combine the prediction frames of the grain categories obtained in step 5.4, suppress some overlapping or incorrect prediction frames through the NMS (non-maximum suppression) method, and perform position regression based on the position coordinates generated in step 5.4 to generate the final prediction frame set, i.e., the grain detection result. The number is determined according to the number of detection frames, where the number of prediction frames for incomplete grains is n, the number of prediction frames for complete grains is m, and the number of prediction frames for all categories is N.

[0071] The grain loss rate S of step 6 is as follows:

[0072] The mathematical model of the grain loss rate S is the ratio of the total number of lost grains to the total number of grains. The grain loss rate is calculated as shown in formula (1).

[0073] S=n / N (1)

[0074] N=m+n

[0075] Where n is the number of incomplete kernels; m is the number of complete kernels; and N is the number of all kernels.

[0076] Step 7: Repeat the test: Repeat steps 1-6 to obtain multiple groups of grain loss rates.

[0077] Step 8: Calculate the average: average multiple groups of data to obtain the final grain loss rate.

[0078] The rice combine harvester grain loss detection method based on SSD-FFNet, first, sampling is carried out at the discharge port of the combine harvester; then, a sieve is used to screen, and the longer straw is screened out, and the grains and small-volume debris are retained; then, the small-volume debris is blown away as much as possible by using a blowing device; then, an industrial camera is used to align with the sampling platform to collect grain images; then, the target detection algorithm SSD-FFNet detects rice grains (including the number of complete grains and incomplete grains), which includes five steps: data preprocessing, feature extraction, prediction frame generation, category prediction and position regression, and prediction frame screening and correction; then, the grain loss rate is calculated according to the number of complete and incomplete grains detected; thereafter, the above steps are repeated to obtain multiple groups of grain loss rates; finally, the multiple groups of data are averaged to obtain the final grain loss rate. The rice combine harvester grain loss detection method includes sampling, screening, blowing, image acquisition and image detection technology, which realizes accurate detection and real-time monitoring of grain loss rate, and provides a reference for the operation status of the rice combine harvester.

[0079] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0080] Although the embodiments disclosed in the present invention are as above, the contents described are only embodiments adopted to facilitate understanding of the present invention and are not intended to limit the present invention. Any technician in the field to which the present invention belongs can make any modifications and changes in the form and details of implementation without departing from the spirit and scope disclosed in the present invention, but the patent protection scope of the present invention shall still be subject to the scope defined in the attached claims.

Claims

1. A rice combine harvester grain loss detection method based on SSD-FFNet, characterized in that: The following steps are included: Real-time sampling of grains at the discharge port of the combine harvester and pre-treatment; Collect images of kernels after screening and peeling; The target detection algorithm SSD-FFNet is used to detect the number of complete and incomplete rice grains in the image, and the grain loss rate is calculated based on the number of grain losses. Repeat the above steps to obtain multiple groups of grain loss rates; average the multiple groups of data to obtain the final grain loss rate.

2. The rice combine harvester grain loss detection method based on SSD-FFNet as claimed in claim 1, characterized in that: The sampled grains are screened with a sieve to remove the long straw and retain the grains and small-volume debris; the small-volume debris is then blown away with a blower.

3. The rice combine harvester grain loss detection method based on SSD-FFNet as claimed in claim 1, characterized in that: The sampling adopts the periodic sampling of the roulette wheel, the sampling period is w minutes, and multiple sampling is carried out in different time periods.

4. The rice combine harvester grain loss detection method based on SSD-FFNet as claimed in claim 1, characterized in that: Normalizing the size of the collected grain image I to obtain a normalized image I1; The normalized image I1 is divided into 6 feature prediction layers for feature extraction to obtain 6 feature maps; Apply prediction boxes of different sizes to each position on the six feature maps; perform convolution on each prediction box to obtain the category score and coordinate offset of each position; The number of complete and incomplete kernels is obtained based on the category score and coordinate offset.

5. The rice combine harvester grain loss detection method based on SSD-FFNet as claimed in claim 4, characterized in that: The grain image I is normalized. Specifically, the grain image I is a pixel matrix, the number of rows of the pixel matrix is ​​the height h1 of the grain image I, and the number of columns of the pixel matrix is ​​the width w1 of the grain image I. The size of the grain image I is normalized to an image I1 of 300×300×3 by upsampling or downsampling.

6. The rice combine harvester grain loss detection method based on SSD-FFNet as claimed in claim 4, characterized in that: Perform feature extraction on the normalized image I1, including: 1) The first part of the feature extraction module: The first feature prediction layer is created through convolution and pooling to extract features. This part consists of thirteen layers, including two convolution layers with a size of 3×3 and 64 convolution kernels, two convolution layers with a size of 3×3 and 128 convolution kernels, three convolution layers with a size of 3×3 and 256 convolution kernels, three convolution layers with a size of 3×3 and 512 convolution kernels, and three maximum pooling layers. The input matrix size of this part is 300×300×3 image I1, and after convolution and pooling operations, the final output size is 38×38×512 feature map y1. 2) The second part: feature extraction module The second feature prediction layer is created through convolution and pooling to extract features. This part consists of a seven-layer network, including three convolution layers with a size of 3×3 and 512 convolution kernels, one convolution layer with a size of 3×3 and 1024 convolution kernels, one convolution layer with a size of 1×1 and 1024 convolution kernels, and two maximum pooling layers. The input matrix size of this part is the feature map y1 of 38×38×512, and after convolution and pooling operations, the final output size is the feature map y2 of 19×19×1024. 3) The third part: feature extraction module The third feature prediction layer is created through convolution and pooling to extract features. This part consists of a two-layer network, including a convolution layer with a size of 1×1 and 256 convolution kernels, and a convolution layer with a size of 3×3 and 512 convolution kernels. The input matrix size of this part is the feature map y2 of 19×19×1024, and after convolution and pooling operations, the final output size is the feature map y3 of 10×10×512. 4) The fourth part: feature extraction module Create a fourth feature prediction layer through convolution and pooling to extract features; This part consists of two branches, which are composed by splicing by adding the number of channels. Branch 1 consists of a two-layer network, including a convolution layer with a size of 1×1 and 128 convolution kernels, and a convolution layer with a size of 3×3 and 128 convolution kernels. Branch 2 consists of an adaptive pooling layer and a convolution layer with a size of 1×1 and 128 convolution kernels. The input matrix size of branch 1 is feature map y3 with a size of 10×10×512, and the input matrix size of branch 2 is feature map y1 with a size of 38×38×512. After convolution, pooling and splicing operations, the final output size is feature map y4 with a size of 5×5×256. 5) The fifth part: feature extraction module Create a fifth feature prediction layer through convolution and pooling to extract features; This part consists of two branches, which are composed by splicing by adding the number of channels. Branch 1 consists of a two-layer network, including a convolution layer with a size of 1×1 and 128 convolution kernels, and a convolution layer with a size of 3×3 and 128 convolution kernels. Branch 2 consists of an adaptive pooling layer and a convolution layer with a size of 1×1 and 128 convolution kernels. The input matrix size of branch 1 is feature map y4 with a size of 5×5×256, and the input matrix size of branch 2 is feature map y2 with a size of 19×19×1024. After convolution, pooling and splicing operations, the final output size is feature map y5 with a size of 3×3×256. 6) Part 6 Feature Extraction Module Create a sixth feature prediction layer through convolution and pooling to extract features; This part consists of two branches, which are composed by splicing according to the number of channels. Branch 1 consists of a two-layer network, including a convolution layer with a size of 1×1 and 128 convolution kernels, and a convolution layer with a size of 3×3 and 128 convolution kernels. Branch 2 consists of an adaptive pooling layer and a convolution layer with a size of 1×1 and 128 convolution kernels. The input matrix size of branch 1 is feature map y5 with a size of 3×3×256, and the input matrix size of branch 2 is feature map y3 with a size of 10×10×512. After convolution and pooling operations, the final output size is feature map y6 with a size of 1×1×256.

7. The rice combine harvester grain loss detection method based on SSD-FFNet as claimed in claim 4, characterized in that: Prediction boxes of different sizes are applied to each position on the six feature maps. Specifically, three prediction boxes with different aspect ratios are constructed on each pixel in the generated feature maps y1, y5, and y6, and the aspect ratios of the prediction boxes include three ratios [1, 2, 1 / 2]; five prediction boxes of different scales and aspect ratios are constructed on each pixel in the generated feature maps y2, y3, and y4, and the aspect ratios of the prediction boxes include five ratios [1, 2, 1 / 2, 3, 1 / 3]; the scales of the prediction boxes generated on the feature maps y1-y6 are [30, 60, 111, 162, 213, 264] respectively.

8. The rice combine harvester grain loss detection method based on SSD-FFNet as claimed in claim 4, characterized in that: Convolution is performed on each prediction box to obtain the category score and coordinate offset of each position. Specifically, the generated prediction boxes are convolved with two convolutional layers of size 3×3. One convolution outputs the category, i.e., the probability scores of complete and incomplete kernels, and the other convolution outputs the position coordinates (x, y, w, h) for regression.

9. The rice combine harvester grain loss detection method based on SSD-FFNet as claimed in claim 1, characterized in that: The category and prediction box of each grain target are obtained according to the category score and coordinate offset; specifically, The obtained grain category prediction frames are gathered together, and the non-maximum suppression (NMS) method is used to suppress a part of the overlapping or incorrect prediction frames. Position regression is performed based on the position coordinates generated above to generate the final prediction frame set, i.e., the grain detection result. The number is determined according to the number of detection frames, where the number of prediction frames for the category of incomplete grains is n, the number of prediction frames for the category of complete grains is m, and the number of prediction frames for all categories is N.

10. The rice combine harvester grain loss detection method based on SSD-FFNet as claimed in claim 1, characterized in that: The mathematical model of the grain loss rate S is the ratio of the total number of lost grains to the total number of grains. The grain loss rate is calculated as shown in formula (1): In the formula, n is the number of incomplete grains; m is the number of complete kernels; N is the number of all kernels.

Images