A Detection and Recognition Method for Cherry Tomatoes Based on Feature Pyramid

Abstract

A method for detecting and recognizing cherry tomatoes based on a feature pyramid, comprising: 1. constructing a backbone network ResNet50 to obtain detailed features of an image by utilizing the strong feature extraction ability of the residual structure; 2. designing a hybrid dilated convolution module to obtain more global features by utilizing the characteristic of dilated convolution to increase the receptive field of features; 3. constructing a feature pyramid to fuse the differential information in features at different levels and eliminate the redundant information generated by the correlation between different features; 4. constructing a detection sub-network to output the probability of a real target and the spatial position offset through a fully connected structure; 5. using the K-means clustering method to obtain the most suitable detection box size; 6. introducing Focal Loss to reduce the problem of imbalance between positive and negative samples, and combining with CIoU Loss to effectively improve the regression accuracy of the bounding box; 7. adopting transfer learning to load the pre-trained weights of the backbone network to accelerate the model convergence speed and optimize the model training effect; 8. using the trained model to detect cherry tomatoes.

Description

Technical Field

[0001] This application relates to the technical field of computer vision image recognition, and more specifically to an object detection method based on a convolutional neural network in the field of deep learning. Background Art

[0002] As one of the main vegetable crops in China, cherry tomatoes are widely planted. The picking of ripe cherry tomatoes requires a large amount of manpower and financial resources, and intelligent picking robots are an effective way to reduce production costs, liberate the labor force and improve work efficiency. The accurate recognition and positioning of fruits and vegetables are one of the prerequisites for constructing an intelligent picking vision system. Therefore, the research on a fast and accurate recognition method for cherry tomatoes has great research significance and application value.

[0003] At present, the cherry tomato detection and recognition methods based on deep neural networks are superior to traditional detection methods in terms of algorithm performance and accuracy. According to the differences in the processing process of detection and recognition, the cherry tomato detection methods based on deep neural networks can be divided into two-stage detection methods and one-stage detection methods. The two-stage method has high accuracy, but the detection speed is slow and the calculation cost is high. The advantage of the one-stage method is that the detection speed is fast, but the accuracy is not as good as that of the two-stage method. Due to the complex growth environment of cherry tomatoes, the fruits are easily blocked by leaves, and the degree of fruit density adhesion is high and the fruits are small, which seriously affects the accuracy of target detection and positioning. Therefore, the research on cherry tomato detection and recognition methods needs to focus on solving the following problems: 1) The problem of missing features of occluded targets; 2) The problem of target detection accuracy and the robustness of the model. Summary of the Invention

[0004] Aiming at the problems existing in the existing deep neural network object detection model in cherry tomato detection, such as low detection accuracy, missed detection and false detection in case of occlusion, etc., the present invention provides a cherry tomato detection and recognition method based on a feature pyramid, which can improve the detection accuracy of cherry tomatoes and achieve real-time detection at the same time.

[0005] The specific steps of a cherry tomato detection and recognition method based on a feature pyramid according to the present invention are as follows:

[0006] Step 1: Construct a backbone network ResNet50. Based on the residual block structure composed of a 1×1 convolution kernel, a 3×3 convolution kernel, and a 1×1 convolution kernel, utilize the strong feature extraction ability of the residual structure to obtain cherry tomato feature maps of different scales. The backbone network is divided into 5 layers. The first layer C1 consists of a 7×7 convolutional layer, a BN normalization layer, and a Mish activation function, and C2 to C5 are respectively composed of 3, 4, 6, and 3 of the above-mentioned residual blocks.

[0007] Step 2: Construct a hybrid dilated convolution component. The C3, C4, and C5 feature maps obtained from the backbone network are subjected to a 1×1 convolution operation with a kernel size to retain the original features while unifying the number of channels to 256. In addition, C3 and C4 will be subjected to 3×1 and 1×3 convolution operations with a kernel size respectively to capture the horizontal and vertical features in the image, so as to enhance the inference ability of the model and improve the problem of target occlusion.

[0008] The feature maps of the above C3, C4, and C5 processed to obtain the same size are subjected to feature fusion, and passed through a hybrid dilated convolution module composed of three 3×3 convolution layers with different dilation rates in series, so as to increase the feature receptive field. To avoid the grid effect of dilated convolution and ensure the integrity of information, the design of this set of dilation rates satisfies the following formula:

[0009] M i =max[M i+1 -2r i ,M i+1 -2(M i+1 -r i ),r i (1)

[0010] where the dilation rates are [r 1 ,...,r i ,...,r n , M i is defined as the maximum distance between two non-zero values in the i-th layer, i ∈ [1, n] and M n =r n . The purpose of formula (1) is to make M 2 ≤K to ensure that there is no grid effect.

[0011] Step 3: Construct a feature pyramid, based on the output feature f obtained by inputting the feature layers C3, C4, and C5 into the hybrid dilated convolution component. The output feature f is subjected to two 1×1 convolution operations with a kernel size to obtain the top layer P7 of the feature pyramid; the output feature f is subjected to a 3×3 convolution operation with a kernel size to obtain the feature pyramid P6; C5’ obtained by subjecting C5 to a 1×1 convolution operation with a kernel size is fused with the output feature f after unifying the number of channels to obtain the feature f1, and f1 is then subjected to a 3×3 convolution operation with a kernel size to obtain the feature pyramid P5; C4’ obtained by subjecting C4 to a 1×1 convolution operation with a kernel size is subjected to feature fusion with the two-fold upsampling result of the feature f1 to obtain f2, and then f2 is subjected to a 3×3 convolution operation with a kernel size to obtain the feature pyramid P4; C3’ obtained by subjecting C3 to a 1×1 convolution operation with a kernel size is subjected to feature fusion with the two-fold upsampling result of the feature f2 to obtain f3, and then f3 is subjected to a 3×3 convolution operation with a kernel size to obtain the feature pyramid P3.

[0012] Step 4: Construct a detection sub-network, which consists of a target classification sub-network and a bounding box regression sub-network. Both sub-networks are fully connected structures composed of 3 convolutional layers with a convolutional kernel size of 3×3, and the output of the feature pyramid is used as the input. The classification sub-network predicts the probability of the existence of real targets at each position for the prior boxes and target categories, and the bounding box regression sub-network regresses the offsets of the 4 spatial positions of each prior box to the nearby real targets.

[0013] Step 5: Use the K-means clustering method to determine the appropriate detection box sizes. The K-means clustering steps are as follows:

[0014] S1. Set 9 clustering centers (i.e., 9 groups of prior box widths and heights).

[0015] S2. Calculate the distance d between the labeled box sizes of the real targets in the cherry tomato image dataset and each clustering center:

[0016] d(box,centroid)=1 - IoU(box,centroid) (2)

[0017] where IoU is the intersection over union of the labeled box and the width and height of the clustering center.

[0018] S3. Assign the labeled boxes to the nearest clustering center, and recalculate the clustering center for each cluster. The calculation method is:

[0019]

[0020]

[0021] where i ∈ [1,9] represents the cluster number, N i represents the number of labeled boxes in the i-th cluster, W i ' represents the average width of all labeled boxes in the i-th cluster, w j represents the width of the j-th labeled box, H i ' represents the average height of all labeled boxes in the i-th cluster, h j represents the width of the j-th labeled box.

[0022] S4. Repeat steps S2 and S3 until the clustering centers tend to fit, and obtain the final 9 sets of appropriate detection box sizes.

[0023] Step 6: Define the loss function for model training. Introduce Focal Loss as the classification loss function of this model. By reducing the loss weight of easy-to-separate negative samples, it focuses on training difficult-to-separate negative samples to solve the influence of extremely imbalanced positive and negative samples during the training process. The Focal Loss calculation steps are as follows:

[0024] LFL (p t ) = -α t (1 - p t ) γ log(p t ) (5)

[0025]

[0026] where \(y\in\{\pm1\}\) represents the true label corresponding to the sample, \(p\in[0,1]\) represents the probability estimate of the model for classifying \(y = 1\), \(\alpha t \in[0,1]\) is a weighting factor used to balance positive and negative samples, \(\gamma\gt0\) is an adjustable focusing parameter, \((1 - p t ) γ is used as a regulating factor to weaken the easily separable negative samples, so that the training focuses on the difficult-to-separate negative samples. The CIoU Loss is introduced as the loss function for bounding box regression prediction. While improving the accuracy of the model for bounding box regression, it also speeds up the convergence rate of the model. To accurately reflect the actual situation between the detection box and the target box, the CIoU Loss incorporates three important geometric factors of the bounding box regression loss: the overlapping area, the distance between the center points, and the aspect ratio. The specific calculation process of the CIoU Loss is as follows:

[0027]

[0028]

[0029]

[0030]

[0031] where \(R CIoU is defined as the penalty term between the detection box and the target box, \(\rho 2 (b, b gt ) is the Euclidean distance between the center points of the two boxes, \(c\) is the diagonal distance of the smallest rectangle enclosing the two boxes, \(v\) measures the difference in the aspect ratio of the two boxes, \(\alpha\) is a positive weighting parameter, \(w gt \), \(h gt \), \(w\), and \(h\) are the width and height values of the true target box and the detection box respectively, and IoU is the intersection over union of the detection box and the true target box. The loss function of this model is:

[0032] L = L FL + \(\beta\)L CIoU (11)

[0033] where \(\beta\) is the coefficient used to balance the classification loss and the bounding box regression loss.

[0034] Step 7: Train the model. Using transfer learning, load the pre-trained weights of ResNet50 to accelerate the convergence speed of the model. The set image input size of the model is 640×640, and normalization processing is performed before inputting into the network. The network parameter update uses the Stochastic Gradient Descent (SGD) method for optimization.

[0035] Step 8: Detect cherry tomatoes with the trained model.

[0036] The innovations of the present invention are as follows:

[0037] 1. The present invention designs a hybrid dilated convolution module. By connecting multiple dilated convolution blocks in series, the generation of holes is effectively avoided. Utilizing the characteristic of dilated convolution to increase the receptive field of features, the feature map integrates more global feature information of the image, enhancing the inference ability and discrimination ability of the model features and reducing the impact caused by incomplete target information.

[0038] 2. The present invention combines Focal Loss and CIoU Loss as the loss function during the training process of the model. Using Focal Loss to reduce the loss weight of easy-to-separate negative samples, thereby focusing on training difficult-to-separate negative samples, and solving the influence of extremely imbalanced positive and negative samples during the training process. CIoU Loss takes into account three important geometric factors: the overlapping area between the real target box and the detection box, the distance between the center points, and the aspect ratio, which can effectively improve the regression accuracy of the model's bounding box.

[0039] The advantages of the present invention are: designed for the target detection task of cherry tomatoes in complex growth environments, effectively improving the target detection accuracy through the above improvement strategies, while achieving real-time target fast detection, providing a relevant basis for intelligent picking technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a schematic diagram of the convolutional neural network model of the method of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0041] The following is a more specific description of the embodiments of the present invention. The flowchart of the embodiments is shown in the appendix Figure 1 The embodiments described by referring to the appendix Figure 1 are only used to explain the present invention and should not be construed as a limitation to the present invention.

[0042] Referring to Figure 1 , the present invention provides a cherry tomato detection and recognition model based on a feature pyramid, which can improve the detection accuracy of cherry tomatoes and reduce the calculation cost.

[0043] A method for detecting and recognizing cherry tomatoes based on a feature pyramid according to the present invention, the specific steps are as follows:

[0044] Step 1: Construct the backbone network. Based on the residual block structure composed of 1×1 convolution, 3×3 convolution, and 1×1 convolution with kernel sizes, utilize the strong feature extraction ability of the residual structure to obtain cherry tomato feature maps of different scales. The backbone network is divided into 5 layers. The first layer C1 consists of a 7×7 convolutional layer, a BN normalization layer, and a Mish activation function. C2 to C5 are respectively composed of 3, 4, 6, and 3 of the above-mentioned residual blocks.

[0045] Step 2: Construct the hybrid dilated convolution component. The C3, C4, and C5 feature maps obtained from the backbone network are subjected to 1×1 convolution operations to retain the original features and unify the number of channels to 256. In addition, C3 and C4 will respectively perform 3×1 and 1×3 convolution operations to capture the horizontal and vertical features in the image, so as to enhance the inference ability of the model and improve the target occlusion problem.

[0046] Fuse the feature maps of the same size obtained after processing the above C3, C4, and C5, and pass them through a hybrid dilated convolution module composed of three 3×3 convolutional layers with dilation rates of dr = 1, 2, and 3 in series, so as to increase the feature receptive field. To avoid the grid effect in dilated convolution and ensure the integrity of information, the design of this set of dilation rates satisfies the following formula:

[0047] M i = max[M i+1 - 2r i , M i+1 - 2(M i+1 - r i ), r i (1)

[0048] where the dilation rates are [r 1 ,..., r i ,..., r n , M i is defined as the maximum distance between two non-zero values in the i-th layer, i ∈ [1, n] and M n = r n . The purpose of formula (1) is to make M 2 ≤ K to ensure that there is no grid effect.

[0049] Step 3: Construct a feature pyramid, based on the feature layers C3, C4, and C5 being input into the hybrid dilated convolution component to obtain the output feature f. The output feature f undergoes two 1×1 convolutional operations to obtain the topmost layer P7 of the feature pyramid; the output feature f undergoes one 3×3 convolutional operation to obtain the feature pyramid P6; C5’ obtained by C5 through a 1×1 convolutional operation is fused with the output feature f after unifying the number of channels to obtain the feature f1, and f1 then undergoes a 3×3 convolutional operation to obtain the feature pyramid P5; C4’ obtained by C4 through a 1×1 convolutional operation is fused with the result of upsampling the feature f1 by two times to obtain f2, and then f2 undergoes a 3×3 convolutional operation to obtain the feature pyramid P4; C3’ obtained by C3 through a 1×1 convolutional operation is fused with the result of upsampling the feature f2 by two times to obtain f3, and then f3 undergoes a 3×3 convolutional operation to obtain the feature pyramid P3.

[0050] Step 4: Construct a detection sub-network, which consists of an object classification sub-network and a bounding box regression sub-network. Both sub-networks are fully connected structures composed of 3 convolutional layers with a convolutional kernel size of 3×3, and the output of the feature pyramid is used as the input. The classification sub-network predicts the prior boxes and the probability of the existence of a real object at each position, and the bounding box regression sub-network will regress the offsets of the 4 spatial positions of each prior box to the nearby real object.

[0051] Step 5: Use the K-means clustering method to determine the appropriate detection box size. The K-means clustering steps are as follows:

[0052] 1. Set 9 clustering centers (i.e., 9 groups of prior box widths and heights).

[0053] 2. Calculate the distance d between the annotation box size of the real object in the cherry tomato image dataset and each clustering center:

[0054] d(box,centroid)=1 - IoU(box,centroid) (2)

[0055] where IoU is the intersection over union of the annotation box and the width and height of the clustering center.

[0056] 3. Assign the annotation boxes to the nearest clustering center, and recalculate the clustering center for each cluster. The calculation method is:

[0057]

[0058]

[0059] where i ∈ [1, 9] represents the cluster number, N i represents the number of annotation boxes in the i-th cluster, and W i' represents the average width of all the labeled bounding boxes in the \(i\)-th cluster, \(w\) j represents the width of the \(j\)-th labeled bounding box, \(H\) i ' represents the average height of all the labeled bounding boxes in the \(i\)-th cluster, \(h\) j represents the width of the \(j\)-th labeled bounding box.

[0060] 4. Repeat steps 2 and 3 until the cluster centers tend to fit, and obtain the final 9 suitable detection box sizes.

[0061] Step 6: Define the loss function for model training. Introduce Focal Loss as the classification loss function of this model. By reducing the loss weight of the easy-to-separate negative samples, it focuses on training the hard-to-separate negative samples to solve the influence of the extremely imbalanced positive and negative samples during the training process. The calculation steps of Focal Loss are as follows:

[0062] \(L\) FL \((p\) t ) = -\(\alpha\) t (1 - p\) t ) γ \(\log(p\) t ) (5)

[0063]

[0064] where \(y\in\{\pm1\}\) represents the true label corresponding to the sample, \(p\in[0,1]\) represents the probability estimate of the model for classifying \(y = 1\), \(\alpha\) t \(\in[0,1]\) is a weighting factor used to balance positive and negative samples, \(\gamma\gt0\) is an adjustable focusing parameter, \((1 - p\) t ) γ is used as a regulatory factor to weaken the easy-to-separate negative samples, so that the training focuses on the hard-to-separate negative samples. Introduce CIoU Loss as the loss function for bounding box regression prediction. While improving the accuracy of the model for bounding box regression, it speeds up the convergence rate of the model. To accurately reflect the actual situation of the detection box and the target box, CIoU Loss takes into account three important geometric factors of the bounding box regression loss: the overlapping area, the distance between the center points, and the aspect ratio. The specific calculation process of CIoU Loss is as follows:

[0065]

[0066]

[0067]

[0068]

[0069] where \(R\) CIoU is defined as the penalty term between the detection box and the target box, \(\rho\)2 (b, b gt ) is the Euclidean distance between the center points of the two bounding boxes, c is the diagonal distance of the smallest rectangle enclosing the two bounding boxes, v measures the difference in the aspect ratios of the two bounding boxes, α is a positive weighting parameter, w gt , h gt , w, and h are the width and height values of the ground truth bounding box and the detected bounding box respectively, and IoU is the intersection over union of the detected bounding box and the ground truth bounding box. The loss function of this model is:

[0070] L = L FL + βL CIoU (11)

[0071] Among them, β is the coefficient for balancing the classification loss and the bounding box regression loss, and β = 2 is set.

[0072] Step 7, Train the model. Using transfer learning, load the pre-trained weights of ResNet50 to accelerate the convergence speed of the model. The set image input size of the model is 640×640, and normalization processing is performed before inputting to the network. The network parameters are updated using the Stochastic Gradient Descent method (SGD) for optimization, where the momentum is set to 0.9, the weight decay is set to 0.0003, the initial learning rate is 0.03, the batch size is 16, and experiments show that the model training converges stably after 50 iterations.

[0073] Step 8, Detect cherry tomatoes with the trained model.

[0074] At the same time, compare the accuracy of the method proposed in this invention with several other classical object detection methods on the same cherry tomato dataset. The results are shown in Table 1 below. The mean average precision (mAP) of the method in this paper can reach 90.12%, which is higher than several other mainstream object detection methods.

[0075] Table 1 Evaluation and comparison of the method in this invention and other object detection methods

[0076]

Claims

1. A detection and recognition method for cherry tomatoes based on a feature pyramid, comprising the following steps: Step 1: Construct a backbone network. Based on the residual block structure composed of 1×1 convolution with kernel size, 3×3 convolution with kernel size, and 1×1 convolution with kernel size, utilize the strong feature extraction ability of the residual structure to obtain cherry tomato feature maps of different scales. The backbone network is divided into 5 layers. The first layer C1 consists of a 7×7 convolutional layer with kernel size, a BN normalization layer, and a Mish activation function. C2 to C5 are respectively composed of 3, 4, 6, and 3 of the above-mentioned residual blocks. Step 2: Construct a hybrid dilated convolution component. The C3, C4, and C5 feature maps obtained from the backbone network are subjected to 1×1 convolution operation with kernel size to retain the original features while unifying the number of channels to 256. In addition, C3 and C4 will respectively perform 3×1 and 1×3 convolution operations with kernel size to capture the horizontal and vertical features in the image, so as to enhance the inference ability of the model and improve the target occlusion problem. Fuse the feature maps of the above-mentioned C3, C4, and C5 processed to obtain the same size, and pass through a hybrid dilated convolution module composed of three 3×3 convolutional layers with different dilation rates in series, so as to increase the feature receptive field. To avoid the grid effect of dilated convolution and ensure the integrity of information, the design of this group of dilation rates satisfies the following formula: M i = max[M i+1 - 2r i , M i+1 - 2(M i+1 - r i ), r i (1) Where, The expansion rates are successively [r 1 ,..., r i ,..., r n , and M i is defined as the maximum distance between two non-zero values in the i-th layer, where i ∈ [1, n] and M n = r n . The purpose of Equation (1) is to make M 2 ≤ K to ensure that there is no grid effect; Step 3: Construct a feature pyramid, based on the feature layers C3, C4, and C5 input into the hybrid dilated convolution component to obtain the output feature f. The output feature f undergoes two 1×1 convolution operations with kernel size to obtain the top layer P7 of the feature pyramid. The output feature f undergoes one 3×3 convolution operation with kernel size to obtain the feature pyramid P6. C5’ obtained by C5 through 1×1 convolution operation with kernel size is fused with the output feature f after unifying the number of channels to obtain the feature f1, and f1 then undergoes one 3×3 convolution operation with kernel size to obtain the feature pyramid P5. C4’ obtained by C4 through 1×1 convolution operation with kernel size is fused with the result of double upsampling of the feature f1 to obtain f2, and then f2 undergoes one 3×3 convolution operation with kernel size to obtain the feature pyramid P4. C3’ obtained by C3 through 1×1 convolution operation with kernel size is fused with the result of double upsampling of the feature f2 to obtain f3, and then f3 undergoes one 3×3 convolution operation with kernel size to obtain the feature pyramid P3. Step 4: Construct a detection sub-network, which consists of an object classification sub-network and a bounding box regression sub-network. Both sub-networks are fully connected structures composed of 3 convolutional layers with a kernel size of 3×3, and use the output of the feature pyramid as the input. The classification sub-network predicts the prior box and the probability that there is a real object at each position for the target category. The bounding box regression sub-network will regress the offsets of the 4 spatial positions of each prior box to the nearby real object. Step 5: Use the K-means clustering method to determine the appropriate detection box size. The K-means clustering steps are as follows: S1. Set 9 clustering centers, that is, 9 groups of prior box widths and heights. S2. Calculate the distance d between the annotation box size of the real object in the cherry tomato image dataset and each clustering center: d(box, centroid) = 1 - IoU(box, centroid) (2) where IoU is the intersection over union of the annotation box and the width and height of the clustering center; S3. Assign the annotation box to the nearest clustering center, and recalculate the clustering center point for each cluster. The calculation method is as follows: Among them, i ∈ [1, 9] represents the cluster number, N i represents the number of bounding boxes labeled in the i-th cluster, W i ' represents the average width of all the bounding boxes in the i-th cluster, w j represents the width of the j-th bounding box, H i ' represents the average height of all the bounding boxes in the i-th cluster, h j represents the width of the j-th bounding box; S4. Repeat steps S2 and S3 until the clustering centers tend to fit, and obtain the final 9 sets of appropriate detection box sizes; Step 6. Define the loss function for model training. Introduce Focal Loss as the classification loss function of this model. By reducing the loss weight of easy-to-separate negative samples, it focuses on training difficult-to-separate negative samples to solve the impact of extremely imbalanced positive and negative samples during training. The calculation steps of Focal Loss are as follows: L FL (p t )=-α t (1 - p t ) γ log(p t ) (5) where, y ∈ {±1} represents the true label corresponding to the sample, p ∈ [0, 1] represents the probability estimate of the model for classifying y = 1, and α t ∈ [0, 1] is a weighting factor used to balance positive and negative samples, γ > 0 is an adjustable focusing parameter, and (1 - p t ) γ is used as a regulation factor to weaken the easy-to-separate negative samples, so that the training focuses on the difficult-to-separate negative samples; CIoULoss is introduced as the loss function for bounding box regression prediction. While improving the accuracy of the model for bounding box regression, it also speeds up the convergence rate of the model; to accurately reflect the actual situation of the detection box and the target box, CIoULoss incorporates three important geometric factors of the bounding box regression loss: the overlapping area, the distance between the center points, and the aspect ratio. The specific calculation process of CIoULoss is as follows: Among them, R CIoU is defined as the penalty term between the detection box and the target box, ρ 2 (b, b gt ) is the Euclidean distance between the center points of the two boxes, c is the diagonal distance of the smallest rectangle enclosing the two boxes, v measures the difference in the aspect ratios of the two boxes, α is a positive weighting parameter, w gt , h gt , w, and h are the width and height values of the true target box and the detection box respectively, IoU is the intersection over union of the detection box and the true target box; the loss function of this model is: L = L FL + βL CIoU (11) where β is the coefficient for balancing the classification loss and the bounding box regression loss; Step 7. Train the model. Adopt transfer learning, load the pre-trained weights of ResNet50 to accelerate the convergence speed of the model; the set image input size of the model is 640×640, and it is normalized before being input into the network. The network parameter update is optimized using the stochastic gradient descent method (SGD); Step 8. Use the trained model to detect cherry tomatoes.

2. A method for detecting and recognizing cherry tomatoes based on a feature pyramid according to claim 1, characterized in that, in the said step 2, the dilation rates of the hybrid dilated convolution components are dr = 1, 2, 3 in sequence, which can make the feature receptive field increase appropriately and effectively improve the model accuracy.

3. A method for detecting and recognizing cherry tomatoes based on a feature pyramid according to claim 1, characterized in that, in the said step 6, the coefficient β for balancing the classification loss and the bounding box regression loss is set to 2, which can make the training of the model in this article reach the optimal state.

4. A method for detecting and recognizing cherry tomatoes based on a feature pyramid according to claim 1, characterized in that, in the said step 7, the network parameter update is optimized using the stochastic gradient descent method (SGD), where the momentum is set to 0.9, the weight decay is set to 0.0003, the initial learning rate is 0.03, the batch size is 16, and the model training in this experiment converges after 50 iterations.

Images