The YOLOv7 (Wang et al., 2023) target detection algorithm, introduced by the original YOLOv4 (Bochkovskiy et al., 2020) research team in July 2022, is a novel and excellent detector that uses instead of efficient aggregation network, the ELAN module that appears in the network structure, to effectively enhance the network learning capability compared to the previous YOLO series. The YOLOv7 network structure mainly consists of the Input layer, Backbone layer, Neck layer and Head layer, Backbone layer, Neck layer, and Head layer. The input layer is the input layer, which scales the input image to a fixed size to meet the input size requirement of Backbone. The backbone layer is the feature extraction layer, and based on YOLOv5, ELAN structure, and MP structure are introduced. Among them, the ELAN structure consists of different convolutional blocks stacked without changing the width-height of the input feature layers and enhances the interaction between each feature layer through expansion, random combination, and splicing to improve the learning ability of the model. The MP structure consists of convolutional blocks of 3x3 size and a Maxpool dual path, which compresses the width-height of the input feature layers to enhance the feature fusion ability of the network. The neck feature fusion network (Neck layer) includes CBS, SPPCSPC, MP, and ELAN, which follow the traditional PAFPN structure to extract three feature layers located in the middle, lower middle, and bottom layers of the backbone part, respectively, to achieve full fusion of multi-scale features. the SPPCSPC structure achieves a full fusion of multi-scale features by introducing a convolutional spatial pyramid (CSP) in the spatial pyramid pool (SPP) structure) structure to improve the perceptual field of the network, while multiple pooling operations are added in parallel in a string of convolutions, using residual edges for optimization and feature extraction. The Head layer uses the anchor mechanism to output feature maps at three scales, large and small, and uses the reparameterized structure RepConv to articulate the regular convolution to adjust the number of channels and prediction, and then the final prediction results are obtained through the processing of CIou loss function and nonlinear maxima suppression (NMS).

Although the YOLOv7 algorithm performs well in common task scenarios (e.g. pedestrian and vehicle detection), there are still many problems in applying it directly to sugarcane stem node recognition in complex environments: for example, in the actual sugarcane environment, dense clusters of sugarcane, small stem node size, and a large number, and a certain degree of occlusion or overlap will lead to serious cases of missed and false detection. To address the above problems, this paper improves the YOLOv7 algorithm in terms of attention mechanism, convolution layer, and loss function to improve the recognition effect in complex environments. The YOLOv7 network structure is shown in Figure 1 .

The attention mechanism refers to the model ignoring irrelevant information and focusing on important information by assigning different weights to the input parts of the network, which can effectively improve the feature extraction ability of the model in complex backgrounds. pyramid feature extraction is used as the backbone network in YOLOv7, and in response to the characteristics of dense sugarcane, small target, large number, and easy to obscure and overlap in the natural environment, this paper adds an attention mechanism to the backbone In this paper, we add an attention mechanism to the backbone network to improve the feature extraction capability and enhance the feature representation capability. The traditional attention mechanisms SE (Hu et al., 2018) (Squeeze-and-Excitation), (Woo et al., 2018) CBAM (Convolutional Block Attention Module), ECA (Wang et al., 2020) (Efficient Channel Attention Module), CA (Hou et al., 2021) (Coordinate Attention), all assign attention weights along channels or spatial locations and require additional sub-network structures and model parameters, which on the one hand cannot generate real 3D weights based on channels or spaces, and on the other hand, the additional sub-network structures inevitably lead to an increase in network complexity. In order to solve the shortage of traditional attention mechanisms, this paper proposes to use SimAM (Yang et al., 2021) non-parametric attention mechanism, whose attention weights are assigned as shown in Figure 2 .

The SimAM attention mechanism generates truly effective 3D weights directly by designing an energy function, without adding additional sub-networks or additional model parameters. The design of the energy function is inspired by neuroscience theory and aims to measure the linear differentiability between neurons to find the important neurons, and for each neuron of the input, the minimum energy function is defined as follows.

where t denotes the target neuron of the input feature in the current channel, where µ,σ ², is the mean and variance of all neurons in the channel to avoid repeated calculations and reduce the computational cost, and λ is the weight constant. To better realize attention, the SimAM module needs to assess the importance of each neuron. In neuroscience, neurons with rich information usually exhibit different firing patterns than surrounding neurons. In addition, activated neurons tend to inhibit surrounding neurons, spatial inhibition, and neurons with spatial inhibition should be given higher importance (Webb et al., 2005). Equation (1) reveals a phenomenon: the lower the energy t of a neuron, the more it differs from surrounding neurons, and therefore the more significant its contribution to visual processing. Thus, the importance of each neuron can be evaluated in terms of e t * . SimAM attention mechanism is added to the ELAN module of the backbone network as well as the Concat connection layer of the neck network, SimAM adjusts the distribution of the attention of the feature map by evaluating the importance of each neuron, so that the channel and the spatial attention work in synergy, the importance of the neurons of the sugarcane stem node will be calculated through the training model, and the information of the secondary disturbances other than the sugarcane stem node will be suppressed, so that the information of the secondary disturbances other than the sugarcane stem node will be suppressed, thus weaken the influence of complex environmental factors on sugarcane stem node recognition. At the same time, SimAM can adaptively adjust the weights of feature mapping and pay more attention to the local area of the target. This can improve the target localization accuracy, reduce the localization error and enhance the feature extraction ability of the backbone network.

The output equation of the attention module is:

The final output is obtained by adding the Sigmoid function to suppress the outliers of the attention weights and performing the dot product operation with the corresponding elements of the input feature matrix.

Aiming at the recognition of sugarcane stem nodes in complex environments and the lack of an attention mechanism in the YOLOv7 network, this paper effectively extracts finer-grained feature information by adding the SimAM attention mechanism to the last layer of the 1×1 CBS of the ELAN module in the Backbone module of YOLOv7 and by incorporating the Concat in the Neck into the SimAM attention mechanism.

In the original YOLOv7 model, ordinary convolutional blocks are used in the feature extraction network Backbone, which mainly consists of a traditional convolutional layer, BN layer, and activation function, and due to the fixed size of the convolutional kernel of the traditional convolutional layer, it has poor robustness to unknown geometric transformations and poor generalization ability. When performing feature extraction, it is difficult for the fixed-size convolution kernel to extract the boundary information of the object accurately because the size and contour of different objects are generally different, which affects the ability of the network to extract the features of the object. The traditional convolutional calculation method is as follows:

where R denotes the size of the partial feature map corresponding to the convolution kernel, W (p_n ) denotes the weight corresponding to the nth sampled point in the sampled region, and X (p ₀ + p_n ) denotes the pixel value size of the nth sampled point. To solve the above problem, a deformable convolutional network (Dai et al., 2017) is introduced in this paper, as shown in Figure 3 :

The network is able to adaptively adjust the size and shape of the convolution kernel for objects of different shapes, which has better robustness and generalization ability compared with traditional convolutional networks, thus enhancing the underlying network’s ability to extract object features. In the variable convolution operator, adding a learnable offset parameter offset to each element in the convolution kernel can make the originally fixed convolution kernel have the ability to adapt to the object shape, and the computational equation is as follows:

In the variable convolution calculation equation (3), Δp_n denotes the offset. Considering the irregularity of the sampling position of the variable convolution, so that the offset is generally fractional, equation (4) is implemented with bilinear interpolation as follows:

where p denotes an arbitrary position in the region corresponding to the deformable convolution, q is the pixel value corresponding to a sampling point in the feature map X, and G(q,p) denotes a two-dimensional bilinear interpolation kernel. In the sugarcane stem node detection task, sugarcane may have complex deformations, such as cane twisting, deformation, or attitude changes, which have an impact on the accurate identification of sugarcane stem nodes. Traditional fixed convolution kernels are difficult to capture these local deformations. Deformable convolution adjusts the sampling position of the convolution kernel by introducing offsets, which allows the convolution operation to better adapt to the target’s deformations and enhances the model’s ability to recognize complex shapes. Deformable convolution utilizes additional convolutional layers to learn the corresponding offsets, and superimposes the obtained offsets on the corresponding pixels in the input feature maps, allowing the convolutional kernel to diverge the sampling in the input feature maps, so that the network can focus on the target function. The embedded deformable convolutional layer can adaptively adjust the sensory field size and position during the convolution process, so that the sampling position around each position during pooling is adaptive and better adapts to the shape and size of the sugarcane stem nodes, thus improving the detection accuracy.

Since deformable convolution adds one more parameter compared to conventional convolution, the 3 × 3 convolution kernel in the ElAN module of the backbone network is replaced with deformable convolution (DCN) in the improved algorithm of this paper, and DCN, BN, and SiLU form the DBS module. And the last layer of 1 × 1 CBS in ELAN is replaced with the SimAM attention mechanism, as shown in Figure 4 , and the replaced structure is represented by DS-ELAN, which enhances the network with a smaller increase in the computational capability of feature extraction and complex background target detection.

The YOLOv7 network uses CIoU (Zheng et al., 2021) Loss as the Bounding Box loss function of the model. This function is mainly used for the regression of the prediction box so that the prediction box of the object is closer to the position of the object labeled bounding box. In the model training, CIoU Loss calculates the distance between the prediction box and the center of the real bounding box, the overlap area, and the aspect ratio of the two boxes for the regression of the bounding box, but it does not take into account the balance of the quality of the training samples, which will lead to the slow convergence and low efficiency of the network, and may result in a worse model due to the random matching of the prediction boxes during training. CIoU is calculated as follows:

Where: b denotes the centroid of the prediction frame; b _gt denotes the centroid of the true frame; ρ represents the Euclidean distance between the two centroids is calculated, and c denotes the diagonal length of the minimum enclosing frame covering the prediction frame and the true frame;α is the balance parameter; v is used to measure whether the aspect ratio is consistent. It can be seen from Equation (7) that when the aspect ratio of the prediction frame and the true value are equal and v takes 0, the penalty term for the aspect ratio in the loss function in CIoU degenerates to 0, resulting in a penalty failure and the final prediction frame cannot fit the true frame. When the prediction frame fits the target frame well, a good loss function should be able to attenuate the penalty of geometric factors.

Because the training data inevitably contains low-quality examples, geometric measures such as distance and aspect ratio can exacerbate the penalty on low-quality examples thus degrading the generalization performance of the model.Wise Iou (Tong et al., 2023) can well improve the sample quality imbalance problem and increase the accuracy of the target detection algorithm. The WIoU loss function is formulated as follows:

R_WIoU range is [1,e), which significantly amplifies the L_IoU of the common quality anchor box.L_IoU range is [0,1]. will significantly reduce the L_IoU of the high-quality anchor frame and the distance between its center of attention when the anchor frame overlaps well with the target frame.

where x y is the coordinate of the center point of the prediction frame, x _gt, y _gt is the coordinate of the center point of the real frame, and Wg, Hg denote the width and height of the minimum enclosing frame. In order to prevent the generation of gradients that hinder convergence, Wg and Hg are separated from the computational graph (the superscript * indicates this operation), which effectively eliminates the factors that hinder convergence, so no new metric, such as aspect ratio, is introduced. Since L _IoU is dynamic, the quality classification criteria of the anchor boxes are also dynamic, which allows WIoU to make a gradient gain allocation strategy that best fits the current situation at each moment.

This strategy reduces the competitiveness of high-quality anchor boxes and also reduces the deleterious gradients generated by low-quality samples. In the sugarcane stem node detection task, some samples may be challenging due to the diversity and complexity of sugarcane images, such as blurred, occluded, or small size of sugarcane images. For these low-quality examples, they may generate noisy or unreliable gradient signals that interfere with the model training process. By introducing category weights through multiple, the WIoU loss function can reduce the weights of low-quality samples relatively, which reduces the impact of these samples on the model parameter updates. Therefore, by reducing the competitiveness of high-quality anchor boxes and reducing the harmful gradient of low-quality samples, the WIoU loss function is able to focus more on the optimization of average-quality anchor boxes and improve the performance of the target detector in general. This strategy helps to make the model more focused on the detection accuracy of important target categories and enhances its ability to handle medium-quality samples, thus improving the effectiveness and performance of overall target detection.

Therefore, this paper uses the WIoU loss function to replace the CIoU in the original network to solve the problem that the regression boxes cannot be matched accurately due to the unbalanced sample quality.

In summary, the improved algorithm adds the SimAM attention mechanism to the feature fusion part of the original YOLOv7 backbone network to enhance the network’s ability to extract target feature information in complex scenes. By using deformable convolution to replace some ordinary convolution layers, the convolution kernel is deformed by more pairs of convolution kernels, which in turn enhances the perceptual field and feature extraction ability of the convolutional network, so that it can better adapt to the shape and position changes of target objects, thus improving the accuracy of object recognition and localization. The CIoU is replaced by WIoU to solve the sample mass balance problem and to make the prediction frame fit the real frame better to improve the detection accuracy. The structure diagram of the improved YOLOv7 network is shown in Figure 5 .

The sugarcane images used in this paper were taken from Su Village, Tao Wei Town, Hengxian County, Guangxi, China, which is a sugarcane plantation. The images were captured by a Xiaomi 10pro digital camera with a resolution of 1080 × 1440 pixels, and the shooting time periods were morning, noon, and afternoon, and a total of 2144 different sugarcane images were obtained by constantly changing the distance and shooting angle. They contain images under uneven conditions of natural scenes such as leaf occlusion, overlapping occlusion, visual similarity to the background image, dense target, backlight, front light, and side light, and are saved in JPG format. The following Figure 6 shows some images taken under different conditions, which were randomly divided into 80% as the training set, 10% as the validation set, and 10% as the test set, forming 1715, 214, and 214 images for model training and testing, respectively. The datasets were then annotated, and the sugarcane stem node bounding boxes in each image were drawn manually using the Roboflow annotation platform, and the annotation files were saved in YOLO text format, which requires the target class, coordinates, height, and width. The process of annotating the sugarcane stem nodes is shown in Figure 7 .

This paper adopted evaluation metrics including precision (P), recall (R), mean average precision (mAP), and F1 score.

P and R refer to the precision and recall of the detection model, respectively. Precision represents the proportion of true positive samples in the samples predicted as positive by the classifier. The recall represents the proportion of true positive samples that are correctly predicted as positive by the classifier among all true positive samples. The formula for calculating precision and recall is:

The F1 score considers both precision and recall, and it can reflect the stability of a model. A higher F1 score indicates a more stable model. The formula for calculating the F1 score is:

mAP is the average precision of each class and the average value of AP, its calculation formula is:

The experimental environment is a 64-bit Ubuntu 22.04 system with an Intel(R) Xeon(R) Platinum 8157 CPU @ 2.30GHz and an NVIDIA GeForce RTX3090 graphics card with 24GB video memory. The study is based on the PyTorch deep learning framework, and the development environment is PyTorch 1.11.0, Cuda 11.3, and Python interpreter version 3.9. The parameters of the experiments conducted in this experiment are shown in Table 1 :

To validate the effectiveness of the improvement points in this paper, five sets of ablation experiments were conducted on the sugarcane dataset using the original YOLOv7 network as a baseline and keeping the environment and parameters uniform. The validation criteria include mAP values and F1 values. The experimental results are shown in Table 2 , where bold font indicates the optimal results in each column and √ indicates the use of the corresponding method.

In order to evaluate the performance of the algorithms, the improved algorithm proposed in this paper is compared with the algorithms such as YOLOv7, YOLOv5, YOLOv8, Faster-R-CNN, and SSD for the detection performance on the dataset, and all the experiments are conducted under the same parameters. The experimental results are shown in Table 3 , the table below, the improved algorithm, the mAP reached 94.53% and the F1 score reached 92.41, which are 3.43% and 2.21 respectively higher than the baseline YOLOv7, compared to the other algorithms in Table 3 with the best overall results. Faster R-CNN and SSD algorithms are affected by the fixed parameters of the anchor frame, which reduces the model detection effect. The YOLOv5 algorithm optimizes the network structure better, and the predicted position regression is more accurate. YOLOv8 has higher accuracy and recall because the C3 structure of the backbone network is replaced by the C2f structure with a richer gradient flow, and the number of channels is adjusted differently for different scale models, which significantly improves the model performance, but it is higher than There is still a certain gap compared with this paper. Comprehensive measurement of different detection algorithms, the improved algorithm in this paper is better, which confirms the effectiveness of the improved algorithm in this paper.