YOLOv8s-CGF: a lightweight model for wheat ear Fusarium head blight detection

Winter wheat FHB image dataset construction

The wheat FHB images were collected in June 2022 at the Rocky Ford FHB nursery, Kansas State University. Four winter wheat varieties “Clark”, “Jagger”, “Overley” and “Everest” were used as the plant materials. The image acquisition device was a high-pixel smartphone, and the shooting process involved top-down photography. The resolution of the images was 3,024 × 4,032, with a total of 231 images. Crop images with a resolution of 832 × 832 from each original image. In this study, data augmentation on the self-built dataset was expanded to 1,386 samples, randomly divided into a training set and a validation set at the ratio of 9:1. An example of wheat FHB images is shown in Fig. 1.

The images were labeled using the LabelImg tool to generate the corresponding XML files. These files contained location information for both healthy and diseased wheat ears, image size, and their respective category information. Mosaic data augmentation was used to handle data samples during training process. This involved randomly cropping and splicing four images into one image as training data to enrich the image background and enhance the model training efficiency.

YOLOv8 model

YOLOv8 is a target detection model and the latest version in the YOLO series of models (Redmon & Farhadi, 2018; Bochkovskiy, Wang & Liao, 2020; Zou et al., 2022). It offers improved speed and accuracy compared to previous YOLO models. YOLOv8 consists of five versions: YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l, and YOLOv8x. These versions are arranged based on model size, with YOLOv8n being the smallest and YOLOv8x being the largest. Since this study focuses on recognizing two categories, namely healthy wheat ears and diseased ears, and considers the need for real-time detection and easy deployment, the YOLOv8s model is selected as the base model. This model has fewer parameters and requires less computational effort.

The YOLOv8s model is made up of four components: Input, Backbone, Neck, and Detection. The Input layer is used to pass the image into the model and perform preprocessing operations on it, while the Backbone network for extracting image features. The Neck layer performs feature fusion on the extracted features, and the Detection layer predicts features of three different dimensions, obtaining category and location information predicted by the network. In this study, the YOLOv8s model served as the base for identifying healthy wheat ears and FHB ears. Additionally, lightweight improvements were made to make it more suitable for deployment on mobile devices.

YOLOv8s-CGF model

Since the FasterNet (Chen et al., 2023) network can effectively reduce model computational redundancy and memory access, improving model speed without affecting accuracy, this article proposes the C-FasterNet module based on the FasterNet module. It serves as the module for the main learning of features in the backbone network of YOLOv8s, which can be used to improve the model speed without affecting accuracy, while reducing the number of parameters and computation. GhostConv (Han et al., 2020) can generate more feature maps from inexpensive operations, fully revealing feature information. Therefore, replacing Conv in the backbone network with GhostConv reduces the model size while improving a certain degree of accuracy. Subsequently, the loss function is replaced with Focal CIoU to speed up model convergence. Finally, the large target detection head in the network is removed for further lightweight. The structure of the YOLOv8s-CGF model is shown in Fig. 2.

C-FasterNet network

C-FasterNet is a proposed module based on FasterNet. The FasterNet network is a new fast neural network proposed in 2023. The design of PConv (partial convolution) in the network exploits the redundancy in the feature maps by applying regular convolution only to some of the input channels, leaving the others unchanged. For consecutive or regular memory accesses, the first or last consecutive channel is considered a representative of the whole feature map for computation. Figure 3 shows how PConv works.

Figure 4 shows the whole structure of FasterNet, which consists of four stages. Each stage is preceded by an embedding stage (a regular Conv 4 × 4 with a step size of 4) or a merging stage (a regular Conv 2 × 2 with a step size of 2) for spatial downsampling and channel count expansion. Each stage has a bunch of FasterNet blocks, and the blocks in the last two stages consume fewer memory accesses while having higher floating point operations per second (FLOPS). Therefore, more FasterNet blocks are placed in the last two stages, and accordingly, more computation is allocated to these stages. The last three layers include global average pooling, Conv 1 × 1, and a fully connected layer for feature transformation and classification.

In this study, based on the above FasterNet module, we propose a C-FasterNet module, which consists of three standard convolutional layers and several FasterNet modules. C-FasterNet is the module that performs the main learning of the residual features. Its structure is divided into two branches: one that uses a stack of several FasterNet modules and a standard convolutional layer, and the other that only uses one standard convolution. Finally, the two branches are subjected to a Concat operation. The C-FasterNet module is used for YOLOv8s network feature extraction, reducing redundant computation and memory access while efficiently extracting spatial features. Its structure is shown in Fig. 5.

GhostConv structure

GhostNet is a lightweight CNN network proposed by the Huawei team. Its core component, the Ghost module (GhostConv), generates a portion of feature maps through original convolution, while the rest is generated using a Cheap operation. This Cheap operation can be a linear transformation of the remaining feature map or a similar feature map generated by Depthwise convolution on the output of the original convolution. The structure of GhostConv is shown in Fig. 6, where C1 and C2 represent the input and output channels, respectively. Half of the output feature maps come from one regular convolution, while the other half is generated by a 5 × 5 Depthwise convolution on the result of the first one. GhostConv achieves the same or even more efficient feature extraction with less complexity than the original convolutional layer. Whereas in CNN networks there is generally a lot of redundant convolutional computation with intermediate feature maps, Ghost convolution forces the network to learn useful features from half of the convolutional kernels. Simultaneously, Depthwise convolution is utilized to generate a feature map for this convolutional portion. The use of a 5 × 5 Depthwise convolution helps widen the receptive field of the generated feature map, enhancing the overall information contained within it.

In this study, GhostConv is employed to replace the convolution module in YOLOv8s, aiming to compress the model rapidly and efficiently and extract target features.

Experimental environment and parameters

All experiments in this study use PyTorch as the deep learning model framework, and the GPU is NVIDIA GeForce RTX3090. The optimizer chosen is the AdamW optimizer, with an initial learning rate of 0.01, momentum set to 0.937, weight decay at 0.0005. The input image size is 640 × 640, the batch size is set to 32, and the model is trained for 200 epochs.

Model evaluation metrics

To evaluate the detection effect of the algorithm on wheat ears, the model performance is examined using recall, precision, average precision and mean average precision as evaluation metrics. The number of parameters, computation, and model size are used to reflect the complexity of the model. The recall reflects the model’s ability to find positive sample targets, the precision reflects the model’s ability to classify samples, and the average precision reflects the model’s overall performance in target detection and classification. The calculation formula is as follows:

(6) $$Precision=\frac{TP}{TP+FP},$$

(7) $$Recall=\frac{TP}{TP+FN},$$

(8) $$AP={\int }_0^{1}P\cdot RdR,$$

(9) $$mAP=\frac{1}{N}{\displaystyle \sum _{i=1}^{N}A}{P}_i,$$where TP denotes the number of correctly identified positive samples, FP denotes the number of incorrectly identified positive samples, FN denotes the number of incorrectly identified negative samples and N represents the number of categories of data. Positive and negative samples are judged by setting the average IoU threshold between the predicted target region and the actual target region, and if the IoU of both exceeds this threshold, it is a positive sample, and vice versa, it is a negative sample. mAP@0.5 is the AP evaluated by the object detection model with an IoU value of 0.5, and mAP@0.5 is the mean of all its classes. mAP@0.5:0.95 represents the average mAP over different IoU thresholds (from 0.5 to 0.95, step size 0.05) (0.5, 0.55, 0.6, 0.65, 0.7, 0.75, 0.8, 0.85, 0.9, 0.95).

Results of ablation experiment

Aiming at the problems that the original YOLOv8s model has a large number of parameters and computations, and the model size takes up a large amount of memory, this study improves the structure of the YOLOv8s network by proposing the C-FasterNet module based on the FasterNet module as the main learning module of the YOLOv8s backbone network. Then, it replaces the Conv in the YOLOv8s backbone network with GhostConv to reduce model size and improve accuracy. GhostConv in the YOLOv8s backbone network to reduce the model size while improving accuracy. Introducing Focal CIoU reduces the loss value and accelerates the model convergence. Finally, removing the large target detection head in the network further reduces the size of the model without affecting the accuracy of the model, making it more suitable for the recognition of wheat FHB and later deployment. To analyze and validate the improved lightweight network model, eight groups of ablation experiment are designed, and the specific experiment results are shown in Table 1.

According to the ablation experiment in Table 1, it can be observed that using the C-FasterNet module to replace the C2f module in the model backbone network reduces the number of parameters and computation amount of the model, and also diminishes the size of the weights generated by the model, while basically keeping the mAP unchanged and slightly lowering the recall rate. The reason for this analysis is that the C-FasterNet module utilizes the FasterNet module, which reduces the number of memory accesses and the model’s computational redundancy, and has higher FLOPS. This results in a reduction of the number of parameters in the model, improvement in model computation speed, and ensures the network is lightweight while maintaining high recognition accuracy. Additionally, from the ablation experiment, it can be observed that after replacing GhostConv, the number of parameters in the model and the computation amount are further reduced, and the model precision is slightly improved. The analyzed reason is that GhostConv can generate more feature maps from cheap operations. Based on a set of intrinsic feature maps, a series of linear transformations are applied at low cost to generate many reshaped feature maps, fully revealing the information of intrinsic features. With the introduction of Focal CIoU, the number of parameters and the amount of computation in the model did not change, and the mAP values remained essentially unchanged. Finally, the large target detection head in the network is removed to further reduce the number of model parameters and the weight size. Figure 7 shows the graph of the loss value changes of the model during the training process, indicating that the Focal CIoU loss function can achieve smaller loss values than CIoU, and the model convergence is faster. Combining the ablation experiment, the YOLOv8s-CGF model proposed in this study ensures high precision in wheat FHB identification and achieves the goal of lightweight.

Comparative analysis of different models

To assess the performance of the improved lightweight network model with other algorithmic models and explore the superiority of the improved algorithms in this study, other target detection algorithms from the YOLO series, such as YOLOv5s, YOLOv6s, and YOLOv7-tiny, are selected for the comparative experiment. The variation of the mAP@0.5 curve for each model on the training dataset is depicted in Fig. 8.

As shown in the figure, mAP improves rapidly in the early stages of training. The model exhibits fluctuations between 10 and 60 epochs, but the overall trend is upward. After 80 epochs, mAP converges and stabilizes, with YOLOv8s-CGF achieving the highest mAP. The performance comparisons for each model are presented in Table 2.

According to the experimental results in Table 2, it can be seen that the YOLOv7-tiny model has the lowest mAP. This is attributed to the YOLOv7-tiny model having fewer parameters and less computation, which hinders achieve higher detection precision. The YOLOv6s model has the highest number of parameters, largest computation and model weight, which does not meet the requirement for lightweight. While the YOLOv5s model achieves good precision with fewer parameters and computation, its precision, recall and mAP are slightly lower compared to the YOLOv8s model. The YOLOv8s-CGF model proposed in this study has the fewest parameters, generates the smallest weight file, and attains the highest precision, recall, and mAP. It achieves superior recognition precision while remaining lightweight, and its detection speed is higher, meeting real-time detection requirements. The recognition results of different models on wheat FHB are shown in Fig. 9.

Statistical analysis

To analyze whether the performance improvement of the modified model is significantly enhanced from a statistical perspective, 10 experiments were conducted for both the pre-improvement and post-improvement models using different random seeds. Paired samples t-test were employed to analyze six performance metrics, and the results are presented in Table 3.

From the table, it is evident that there is a significant difference among groups in terms of the precision metric (P < 0.05). Therefore, it is considered that the improvement method has a significant impact on precision (i.e., significantly improving the model’s precision). On the recall metric, there is no significant difference among groups (P > 0.05), indicating that the improvement method has no significant impact on recall (i.e., cannot significantly improve the model’s recall). Regarding the mAP@0.5:0.95 metric, there is a significant difference among groups (P < 0.05), suggesting that the improvement method has a significant impact on mAP@0.5:0.95 (i.e., significantly improving the model’s mAP@0.5:0.95). Due to the fixed parameters, computation, and model size before and after the model improvement, and the significant reduction in parameters, computation, and model size after the improvement, it is considered that the improved model exhibits significance in these three metrics compared to the original model. In summary, the improved model shows significant differences in all five metrics, indicating that the proposed improvement method in this study significantly enhances model performance.