YOLOv8 is a state-of-the-art (SOTA) model that builds upon the successes of previous YOLO versions, incorporating novel features and enhancements to further improve performance and versatility. Specific innovations include a new backbone network, a new Anchor-Free detection head, and a novel loss function. Anchor-Free Detection Head: Traditional object detection models utilize anchor boxes to determine the position and size of targets. In contrast, the Anchor-Free detection head learns the key-points or bounding boxes of the targets, thus eliminating the need for anchor boxes. This approach enables the model to better adapt to targets of varying sizes and shapes while reducing the complexity associated with tuning anchor boxes. Novel Loss Function: The loss function serves as feedback during training, assisting the model in fine-tuning its parameters for better target approximation. Currently, the YOLOv8 series has introduced five different versions, namely YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l, and YOLOv8x. The model’s parameter and computational complexity increase with the depth and width of the model. Users can choose the appropriate network structure based on their application scenarios. The YOLOv8s version employs a lighter network structure and fewer training data, aiming to maintain relatively fast detection speed and high accuracy while efficiently deploying on embedded devices and small applications (Wang G. et al., 2023). This makes YOLOv8s an ideal choice for real-time object detection applications. Therefore, this paper adopts the YOLOv8s model to meet the demand for efficient object detection. The YOLOv8 model detection network structure, as illustrated in Figure 2 below, comprises the Backbone, FPN, and Head.

The Backbone serves as YOLOv8’s primary feature extraction network. Images fed into this network initially undergo feature extraction to produce what is commonly referred to as feature layers, a comprehensive set of features derived from the input images. These Feature Pyramid Network (FPN) in YOLOv8 is an augmented feature extraction component. Three significant feature layers obtained from the backbone network are further integrated in this section. The objective of this feature fusion is to combine feature information from various scales. The FPN continues to extract features from the already obtained significant feature layers. YOLOv8 still employs the Panet architecture, which not only up-samples the features for fusion but also down-samples them for an additional fusion. The Head in YOLOv8 serves as the classifier and regressor. Through the Backbone and FPN, we can obtain three enhanced, significant feature layers.

The network structure of the cotton pest and disease identification model based on CFNet-VoV-GCSP-LSKNet-YOLOv8s proposed in this paper is illustrated in Figure 3.

YOLOv8 is a state-of-the-art (SOTA) model, further categorized into YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l, and YOLOv8x. The YOLOv8 detection network structure consists of Backbone, FPN, and Head. While the Backbone has a large number of parameters and a long training time, it falls short in the detection of small objects. To address these limitations, we propose a cotton pest and disease identification model based on CFNet-VoV-GCSP-LSKNet-YOLOv8s. Firstly, the CFNet module replaces all C2F modules in the YOLOv8s Backbone. Then, a feature integration operation is inserted in the Backbone, effectively utilizing a large proportion of the Backbone to fuse multi-scale features, thereby improving the model’s recognition rate. Secondly, in the Head of YOLOv8s, the VoV-GCSP module replaces all C2F modules, enhancing the features extracted by the Backbone while also reducing the model size without sacrificing accuracy. Additionally, the LSKNet attention mechanism is incorporated into both the Backbone and Head to improve the detection of small objects. Lastly, the XIoU loss function is introduced to enhance model convergence, thereby achieving accurate identification of cotton pests and diseases.

In the YOLOv8 Backbone, the C2F module attempts to fuse shallow feature maps with high resolution but limited semantic information with deep feature maps that have low resolution but rich semantic content. However, we argue that this approach might be insufficient for effective multi-scale feature fusion, especially when compared to heavy classification backbones where the parameters allocated for feature fusion are limited. To address this issue, we propose a new architecture named Cascaded Fusion Network (CFNet). Apart from the initial high-resolution feature-extracting Backbone and several blocks, we introduce multiple cascading stages to generate multi-scale features within CFNet. Each stage consists of a sub-backbone for feature extraction and an extremely lightweight transformation block for feature integration. This design allows for a more in-depth and effective fusion of features, leveraging a large proportion of the Backbone’s parameters. By replacing all C2F modules in the Backbone with CFNet and then inserting feature integration operations, we achieve effective fusion of multi-scale features across a significant portion of the Backbone.

The core design philosophy of CFNet involves introducing multiple cascading stages, each stage consisting of a specialized feature extraction sub-backbone and an extremely lightweight transformation block, effectively capturing and merging multi-scale features from fine-grained to coarse-grained. These cascading stages not only process the output from the previous stage but also deeply interact with the corresponding features of the main backbone network, achieving complex feature integration. The feature maps produced by each cascading stage are optimized and merged through specifically designed transformation blocks, enhancing the model’s ability to represent features. This structure is especially suitable for object detection tasks that require efficient multi-scale feature fusion. By optimizing and deeply integrating features at different levels, CFNet improves model performance while maintaining relatively low computational costs.

Suppose X i represents the output feature map of the ith cascading stage of the input image, where i represents the sequence number of the cascading stage. F represents the feature extraction function, and T represents the feature transformation function (lightweight transformation block). Thus, each cascading stage can be formally represented as shown in Equation 1:

In CFNet, M represents the total number of cascading stages. X 0 is the initial high-resolution feature map (as shown in Equation 2). For feature fusion, suppose P j denotes the j fused feature map, corresponding to different spatial resolutions, such as P 3 , P 4 , P 5 etc. Each P j can be calculated through the output of the cascade and the corresponding transformation function T j .

In CFNet,. refers to the output of the final cascading stage. The network architecture of CFNet is illustrated in Figure 4.

The CFNet architecture commences by inputting an image with spatial dimensions of H × W through a neck and N successive blocks, extracting high-resolution features with dimensions of H 4 × W 4 . These features are subsequently directed into M cascaded stages for the extraction of multi-scale features. Prior to entry into the M cascaded stages, the extracted high-resolution features are downscaled using a 2×2 convolution kernel with a stride of 2. The network’s architecture at each stage maintains a consistent structural format but varies in scale, comprising differing numbers of processing blocks. Each stage consists of a sub-backbone network and an ultra-lightweight transition block, both dedicated to the extraction and integration of features. For clarity, the assemblage of blocks within each stage, addressing features of the same scale, is termed a block group. The three block groups within the ith stage encompass n i 1 , n i 2 , and n i 3 blocks, respectively. In the final block group of each stage, a so-called focal block is implemented to enhance feature processing. Each stage outputs features P3, P4, P5 with strides of 8, 16, 32, respectively, of which only P3 features are utilized for input into the subsequent stage. In the network’s final stage, features P3, P4, and P5 are amalgamated, serving dense prediction tasks. By substituting all C2F modules in the backbone with CFNet and incorporating feature integration operations, effective fusion of multi-scale features is achieved throughout a significant portion of the backbone. Figures 5 and 6 provide more details about transition blocks and focus blocks.

As depicted in Figure 5, When given C₃, C₄, and C₅ as inputs, the transition block produces outputs P₃ and P₄. The term “Conv -dx” refers to a 1×1 convolution operation that outputs a channel number of dx, where dx matches the channel number of the input feature C_x. The circles marked with “+” and “C” represent element-wise addition and concatenation operations, respectively. Additionally, the notation “2x” is used to indicate the upsampling of features by a scaling factor of 2.

This design facilitates the effective integration of multi-scale features. By adjusting the channel dimensions through 1×1 convolutions and managing the spatial resolutions via addition, concatenation, and upsampling operations, the transition block efficiently processes the varied scales of the input features (C₃, C₄, C₅) and transforms them into the desired output formats (P₃, P₄), which are then suitable for subsequent stages of the network’s processing pipeline.

As shown in Figure 6, N is the number of channels of the output feature. d7×7 represents the 7×7 depth convolution, a7×7 represents the window size, and R is the expansion rate of the additional convolution. GELU is the activation function. Each d7×7 or a7×7 is followed by a LayerNorm layer and a GELU unit. This paper proposes a novel focus block to enlarge the receptive fields of neurons in the last block group of each stage as an effective alternative strategy. The design of the focus module introduces extended depth convolution and two skip connections in the ConvNeXt module, thereby achieving the integration of fine-grained local interaction and coarse-grained global interaction.

The YOLOv8 network employs a substantial number of C2F modules in its neck for feature extraction. However, this structure results in an increase in computational complexity and the number of parameters, leading to significant time consumption. Lightweight networks like Exception and ShuffleNet address the time-consuming issue of standard convolutions by utilizing depth-wise separable convolutions (DSC), albeit at the cost of sacrificing accuracy. The GSConv convolution module is an innovative approach that combines Standard Convolution (SC), Depth-Wise Convolution (DWConv), and channel shuffle operations. The core idea involves partitioning the input channels into multiple groups, performing independent depth-wise separable convolutions on each group to reduce computational complexity. This design aims to mitigate the issue of low recognition accuracy due to insufficient feature extraction and fusion capabilities. The groups are then recombined through channel shuffling. GSConv combines SC, DSC, and Shuffle, exhibiting performance similar to SC but with lower computational costs. The depth layer calculation is shown in Equation 3, and the GSConv layer calculation is shown in Equation 4.

W and H represent the width and height of the feature map, respectively, and m×n is the size of the convolution kernel. Pin and Pout represent input and output function channel numbers. In scenarios where the input feature channel count escalates, the computational demand of the GSConv convolution diminishes, yet it retains a feature extraction proficiency analogous to its contemporaries. The integration of GSConv has been instrumental in the strategic simplification of the model’s complexity. To augment the inference velocity of the network model, while concurrently preserving its precision in detection, we have implemented the VoV-GSCSP module, building upon the foundational GSConv module. The VoV-GSCSP represents a sophisticated hybrid network architecture, which skillfully merges the attributes of GSConv with the essence of VoVNet, supplemented by the incorporation of (Squeeze-and-Excitation, SE) blocks. This architectural design is meticulously tailored to enhance both the quality and efficiency of feature extraction. By segmenting the convolutional layers into discrete groups, the Grouped Separable Convolution effectively minimizes the parameter count and computational complexity. The Squeeze-and-Excitation blocks intensify the network’s representational prowess by concentrating on salient channel features. This innovative structural design endows the VoV-GSCSP module with the capability to sustain high computational efficiency while simultaneously elevating the feature representation and overall performance of the network. The module, engineered with a one-off aggregation methodology, markedly amplifies the inference speed of the network model, all the while maintaining its superior detection accuracy. The configurations of the GSConv convolution and the VoV-GCSP network are exemplified in Figure 7 and Figure 8.

Attention mechanisms serve as a straight forward effective approach yet to enhance neural representations. Channel attention modules like SE blocks utilize global average information to re-weight feature channels, while spatial attention modules such as GENet, GCNet, and SGE enhance the network’s capability to model contextual information via spatial masks. Techniques like CBAM and BAM amalgamate channel and spatial attentions, leveraging the strengths of both. Beyond channel/spatial attention mechanisms, kernel selection is another adaptive and effective technique for dynamic contextual modeling. LSKNet is designed based on attention mechanisms and kernel selection technologies to better model the features of different targets in remote sensing scenarios. It also boasts advantages like relatively fewer parameters and computational complexity, thereby facilitating improved computational efficiency and speed in practical applications. LSKNet is a novel neural network architecture specifically aimed at remote sensing object detection tasks. It enhances contextual modeling and feature extraction through selective mechanisms and adaptive spatial aggregation, consequently improving the performance in small object detection. A detailed structural comparison is shown in Figure 9.

The Loss Function is a metric that measures the difference between the predicted values of a model and the actual values. During training, the model attempts to minimize the value of the loss function to improve its accuracy. YOLOv8s adopts the CIoU loss function, composed of position, confidence, and class functions. This traditional loss function generally relies on the aggregation of bounding box regression indicators, without considering the mismatch in direction between the required ground truth boxes and predicted boxes, leading to slow convergence and low efficiency. The XIoU loss function plays a crucial role in object detection tasks by emphasizing varying degrees of overlap between targets. By combining the regression of predicted boxes with real boxes, this loss function prevents issues such as overlapping center points and identical aspect ratios that would degrade into the IOU loss function. This ensures the effective completion of boundary box regression, improving the robustness of the bounding boxes. Therefore, in this study, the XIoU loss function is introduced as an improvement to the model. Compared to the original CIoU loss function, the penalty term gradient of XIoU is smoother, resulting in smaller regression errors and better regression performance. It also effectively enhances the recognition accuracy of cotton leaf diseases and pests.

XIoU calculation formula is as shown in Equation 5:

The penalty term is defined as shown in Equation 8:

As shown in Equations 5–8, IoU stands for the traditional regression loss. ρ 2 represents the squared Euclidean distance between the two rectangular bounding boxes. c 2 represents the square of the diagonal distance between two rectangular boxes. b and b^gt denote the central points of the two bounding boxes. α weight coefficient.v is used to measure the consistency of the relative proportions between the two boxes. w^gt , h^gt , w and h are the width and height of the two boxes, respectively. The primary goal of XIoU is to improve the IoU metric by considering the intersection area between the boxes, offering a better representation of their overlap. The parameter α is used to adjust the difference between XIoU and IoU, thereby reflecting the similarity between the boxes more accurately and accelerating the network’s convergence.

The model was trained using the PyTorch framework on a laboratory server equipped with an Intel Core i9-10900KF processor, 16 GB of CPU memory, and an NVIDIA GeForce RTX 3080 GPU. The operating environment was Windows 10, with Python 3.8, PyTorch 1.11.0, and CUDA 13.0 used for algorithmic optimization. Training parameters included 150 epochs, a batch size of 8, and an image input resolution of 640×640 pixels. All other settings were kept at their default values.

Performance metrics used for model evaluation included Precision (P), Recall (R), Mean Average Precision (mAP), and model size. Precision is defined as the fraction of true positives among the predicted positives, while Recall measures the fraction of actual positives correctly identified by the model. Mean Average Precision (mAP) serves as a comprehensive performance metric. The above indicators such as Equations 9–12 shown.

In the equations, TP represents the number of true positives, FP stands for false positives, and FN signifies false negatives.

To validate the superior performance of the proposed CFNet- VoV-GCSP -LSKNet -YOLOv8 architecture for the identification of six types of cotton pests and diseases, we compared our model with the original YOLOv8s algorithm, as shown in Table 2.

From Table 2, it can be observed that the method proposed in this paper for identifying six types of cotton pests and diseases—namely, noctuid larvae, cotton angular leaf spot, cotton boll rot, cotton powdery mildew, healthy cotton, and cotton black spot—achieves an average precision mean (mAP@0.5) improvement compared to the original model of 0.9%, 0.6%, 0.1%, 2.4%, 1.1%, and 2.2%, respectively.

Among the six types of cotton pests and diseases, the average precision mean (mAP@0.5) for cotton black spot is the lowest, with only 74.3%. Analysis indicates that the blurriness of the original data images led to this subpar performance. However, with the application of our method, there is a 2.2% improvement over the original model, resulting in an overall average precision mean (mAP@0.5) of 93.7%, an increase of 1.2% compared to YOLOv8s. This demonstrates that our model’s feature extraction capability has been enhanced for images with suboptimal quality. The results of the method proposed in this paper for identifying cotton pests and diseases are shown in Figure 10:

To validate the efficacy of the improvements made to the original algorithm by the cotton pest and disease identification method based on the CFNet-VoV-GCSP-LSKNet-YOLOv8s network structure proposed in this paper, an ablation study was designed. The original network and the network improved with various modules were tested on a test dataset. The results are shown in Table 3:

To further validate the effectiveness and practicality of the method proposed in this paper, the location loss values of CFNet-VoV-GCSP-LSKNet-YOLOv8s and YOLOv8s are shown in Figure 11 after 150 training iterations. As can be seen from Figure 11, the convergence speed of the proposed method is faster, and its convergence performance is superior to that of the YOLOv8s model.

To verify the effectiveness of the cotton pest and disease identification method based on the CFNet-VoV-GCSP-LSKNet-YOLOv8s model proposed in this paper, we compared it with YOLO v5s (Jiang et al., 2022), YOLOX (Ge et al., 2021; Zhang et al., 2022b), YOLOv7 (Cao et al., 2023; Wang C-Y. et al., 2023), Faster R-CNN (Li, 2021; Qiao et al., 2021).

YOLO v8s, YOLOv8n, YOLOv7-tiny, CenterNet and EfficientDet. Among them, YOLOv8s, YOLO v5s, YOLOv7, YOLOv8n, YOLOv7-tiny, CenterNet and EfficientDet are currently mainstream object detection algorithms, while YOLO X and Faster R-CNN have shown better performance in other studies. To validate the superiority of the proposed method, all model training processes maintained consistent parameter settings. The comparison results are shown in Table 4.

As delineated in Table 4, the proposed method demonstrates marked improvements in performance metrics over existing methods. Specifically: When compared with YOLO v5s, YOLO v7, YOLOv8n, YOLOv7-tiny, EfficientDet and YOLOv8s, the precision of our model improved by 1.4%, 11.4%, 5.5%, 2.2%, 2.1% and 2% respectively. Additionally, recall rates saw increases of 1.9%, 11.6%, 5.4%, 49.2%, 20% and 1%, and the mean average precision (mAP) advanced by 2.3%, 12.5%, 3.7%, 21.8%, 12.1% and 1.2%. While our method registers a slight decline in precision relative to YOLOX, it compensates with a 4% increase in recall and a 5% boost in mAP. As for Faster R-CNN, although the recall rate was marginally lower by 0.5%, the model achieved a substantial enhancement in precision by 29.1%, along with a 9% improvement in mAP. Compared with CenterNet, although the precision is 5.9% lower, the mAP is 4.9% higher and the recall rate is 25.6% higher. In terms of computational resource consumption, our model is second only to YOLO v5s, YOLOv8n and YOLO v8s. The detection time of the improved algorithm is 8.01ms. Although slightly lower than the fastest detection speed YOLOv8n and YOLO v8s, other performance indicators of the detection algorithm are better than this model. Therefore, based on the overall detection performance indicators of the model, the algorithm in this paper has great advantages in both recognition accuracy and speed. Collectively, these results validate the effectiveness of the proposed method, positioning it as superior in object detection performance. Figure 12 offers further insights into the comparative performance of various models. While the proposed method slightly lags behind YOLO v5s in terms of convergence speed during the initial 17 iterations, it surpasses all competing models in both convergence speed and mAP following the 17th iteration.