During the infection process of cucumber leaves by downy mildew, the primary physiological characteristicwas the appearance of dark brown lesions on the leaf surface, a phenomenon caused by Pseudoperonospora cubensis^[21], as shown in Fig. 1. This pathogen spreads rapidly and caused significant damage, severely affecting cucumber production and quality. The development and spread of conidiophores and sporangium were closely related to the occurrence of cucumber downy mildew. Sporangium, which growed at the tips of conidiophores, was the fundamental structure responsible for the propagation of the downy mildew pathogen. Their primary function was to produce and release spores under favorable environmental conditions. These spores can survive at lower temperatures and were released when conditions were suitable. They were then spread to healthy plants by wind and rain, entering the plant through stomata and initiating a new infection cycle^[22-23]. Conidiophores, which growed from the underside of the leaf surface, developed under suitable environmental conditions, such as warmth and humidity, thereby increasing the number of sporangium^[24]. The length and position of conidiophores had a decisive impact on the spore dispersal range, determining how far the wind and the dispersion pattern upon landing can carry the spores. Therefore, the precise identification of sporangium and conidiophores was not only crucial for a deeper understanding of the biological characteristics and dissemination mechanisms of the disease but also enabled researchers to more accurately predict the occurrence and spread of the disease, thereby advancing the control measures for cucumber downy mildew.

The cucumber downy mildew leaf samples were collected from the Modern Agricultural Science and Technology Innovation Base of the Tianjin Academy of Agricultural Sciences and the Plant Protection Institute of the Tianjin Academy of Agricultural Sciences. To further study the distribution of conidiophores and sporangium of cucumber downy mildew, microscopic images of the pathogen in the experimental samples were obtained. The experiment began with the preparation of three 1.5 mL centrifuge tubes, each containing 1 mL of distilled water. The downy mildew pathogens on the leaf surface were gently brushed into the centrifuge tubes using a soft brush, creating a suspension containing conidiophores and sporangium. To accurately analyze the distribution of sporangium and conidiophores, an Olympus BX51 microscope set to 200x magnification was used to capture microscopic images by extracting suspension samples from the three different centrifuge tubes. Each microscopic image had a resolution of 1 280 pixels × 960 pixels. To ensure the representativeness and reliability of the data, each centrifuge tube sample was used to generate 550 independent microscopic images, resulting in a total of 1 650 images. These images were stored in JPG format to facilitate subsequent image analysis and data processing. The collected microscopic images were shown in Fig. 2. The images were divided into training, validation, and test sets in a ratio of 8∶1∶1.

The entire experiment was conducted on a Windows 10 (64-bit) operating system, with a computer equipped with 64 GB of memory, an Intel (R) Core(TM)i9-13900KF @ 3.0 GHz CPU, and an NVIDIA GeForce RTX 4090 GPU with 24 GB of VRAM. The deep learning framework used was PyTorch 1.13.1, with CUDA version 11.7. During the model training process, the number of iterations (epochs) was set to 150, and the batch size for each training session was set to 32. The Adam optimizer was used for learning rate adjustment, while the other parameters during the training process were kept at their default settings.

Evaluation metricswere crucial for assessing the effectiveness of object detection algorithms. The improved YOLO v8s-OBB model used precision (P), recall (R), and mean Average Precision (mAP) as the evaluation metrics for comparing its performance with other models.

The CBAM attention mechanism^[25] mainly consisted of two sequential sub-modules: the channel attention module (CAM) and the spatial attention module (SAM). Its basic structure was shown in Fig. 4 CBAM sequentially integrated attention weights along the channel and spatial dimensions based on the input feature map and multiplied them with the input features to generate new features, which was beneficial for extracting meaningful information from the feature map. The introduction of the CBAM attention mechanism was illustrated in Fig. 3. CBAM was chosen over other attention mechanisms, such as SE or Non-local modules, due to its lightweight design and superior performance in enhancing the model's ability to focus on critical regions without significantly increasing computational overhead. CBAM effectively balanced detection accuracy and computational efficiency, making it especially suitable for detecting small and rotated objects like sporangia and conidiophores in microscopic images. Furthermore, CBAM's modular design ensured easy integration into existing deep learning architectures without introducing excessive complexity.

In this model, CBAM was inserted after the C2f module. The C2f module extracted and fused features from the input data. Still, it may need to fully emphasize small target features, critical for detecting minute objects such as sporangia and conidiophores. By introducing CBAM at this stage, the attention mechanism optimized feature representation and enhanced the model's ability to focus on the most critical regions in microscopic images. Specifically, after the C2f module processed and fused the features, CBAM applied channel and spatial attention to prioritize the areas most relevant for target detection, thus enabling the accurate detection of small and rotated objects in microscopic images. This strategic placement ensured that the model can handle complex backgrounds and overlapping targets effectively while maintaining high precision and recall. Following the CBAM module, the model proceeded to the lightweight shared convolutional detection head (LSCD), further improving multi-scale feature detection.

CBAM calculated the attention map for a feature map in two dimensions: channel attention and spatial attention. For an intermediate layer feature map, denoted as F ∈ R^C×H×W, CBAM inferred a one-dimensional channel attention map and a two-dimensional spatial attention map. The calculation process was as follows:where symbol ⊗ represented element-wise multiplication. The channel attention was multiplied by the input feature map F′, and then the spatial attention of F′ was calculated. Both of the above were multiplied to obtain the final output F″.

Channel attention mainly focused on which parts of the input microscopic images contained the conidiophores and sporangium of cucumber downy mildew. To effectively calculate the channel features, max pooling, and average pooling were used to compress the feature map along the spatial dimension, resulting in and . Multi-layer fully connected network was then used to process these two different descriptive directions to calculate the channel attention, as shown in equations (3) and (4):where W₀ ∈ R^C/r×C and W₁ ∈ R^C×C/r, the Sigmoid function was used as the activation function for W₀.

The spatial attention mainly focused on the positional information of the input image features. Firstly, max pooling and average pooling were applied along the channel dimension to obtain two different feature descriptors, and . These two features were then combined through an operation, and a convolution operation was used to generate the spatial attention feature map M_S (F) ∈ R_H×W. The calculation process was as follows:where f^7×7 represented the convolutional layer with a 7×7 convolution kernel.

The lightweight shared convolutional detection head (LSCD) was an efficient and resource-friendly component of object detection models^[26], as shown in Fig. 5. LSCD's design philosophy involved optimizing the detection performance for targets of different sizes through shared convolution operations while simultaneously reducing the demand for computational resources. LSCD can integrate multi-scale features, enabling the model to more accurately identify targets of various sizes.

The original YOLO v8s-OBB model was improved to accurately detect the rotated and tilted sporangium and conidiophores of cucumber downy mildew in microscopic images. Adopting the LSCD, which integrated multi-scale features and shared convolution operations, optimized the detection performance for targets of different sizes. Meanwhile, by implementing 1×1 and 3×3 convolution operations on the P3, P4, and P5 feature layers of various scales, the model effectively reduced the demand for computational resources while improving target detection accuracy. In addition, each scale layer was equipped with specific regression and classification layers and a Conv_OBB layer explicit for oriented bounding boxes to ensure accurate prediction of the targets' shapes and orientations. By integrating LSCD into the detection head, the performance of the YOLO v8s-OBB model in detecting cucumber downy mildew sporangium was significantly enhanced, achieving high-precision detection of small targets with specific morphologies.

To achieve accurate detection of sporangium and conidiophores of cucumber downy mildew in microscopic images, several advanced rotational target detection models were compared, and the resultswere shown in Table 1. According to the results in Table 1, the YOLO v8s-OBB model demonstrated outstanding performance across multiple key metrics. Compared with other models, YOLO v8s-OBB had the highest precision, reaching 94.9%, which was higher than YOLO v7s-OBB's 94.6% and YOLO v5s-OBB's 94.1%, with improvements of 0.3 percentage points and 0.8 percentage points, respectively. In contrast, the precision of S2ANet, H2RBox, and R2CNN was 91.2%, 93.1%, and 90.2%, respectively, making YOLO v8s-OBB higher by 3.7 percentage points, 1.8 percentage points, and 4.7 percentage points compared with that of these models. Regarding recall, YOLO v8s-OBB's score of 85.9% was slightly lower than H2Rbox's 87.2%, but it was 0.8 percentage points and 1.6 percentage points higher than that of S2ANet and R2CNN, respectively. For mean average precision (mAP@0.5), YOLO v8s-OBB reached 95.3%, matching YOLO v7s-OBB but surpassing YOLO v5s-OBB by 0.3 percentage points, and it was significantly higher than S2ANet's 92.3% and R2CNN's 91.1%. Regarding mAP@0.5: 0.95, YOLO v8s-OBB achieved 69.1%, slightly higher than YOLO v7s-OBB's 68.7% and YOLO v5s-OBB's 68.3% and better than S2ANet's 65.1% and R2CNN's 64.2%. Regarding computational complexity, YOLO v8s-OBB had 2.96 × 10¹⁰ FLOPs, significantly lower than S2ANet's 1.31 × 10¹¹ FLOPs and R2CNN's 1.21 × 10¹¹ FLOPs, demonstrating greater computational efficiency. Additionally, the YOLO v8s-OBB model size was only 11.4 MB, the smallest among all models, making it suitable for environments with limited storage space. These advantages mainly stemed from the efficient architecture design and optimized algorithms of YOLO v8s-OBB, which significantly reduced computational complexity and model size while maintaining high precision. In contrast, S2ANet and R2CNN had higher computational resource consumption, while the YOLO series models achieved a balance of performance and efficiency through network structure and loss function optimization. Considering precision, recall, mean precision, computational complexity, and model size, YOLO v8s-OBB was the best-performing model, making it a suitable base model for further improvements and optimizations. Its balanced accuracy and resource efficiency performance make it an ideal choice for applications in real-time and resource-constrained environments.

Ablation experiment was conducted to further evaluate the impact of differentnetwork structures of YOLO v8-OBB on the detection accuracy of sporangium and conidiophores in microscopic images. As shown in Table 2, the YOLO v8-OBB model exhibited different performance characteristics across various network structures (n, s, m, l, x). YOLO v8s-OBB performed best in terms of precision, achieving a value of 94.9%, which surpassed other structures, notably exceeding the precision of YOLO v8m-OBB at 88.4% and YOLO v8l-OBB at 91.1%. Although YOLO v8x-OBB slightly led in mAP@0.5 with a score of 95.5%, YOLO v8s-OBB follows closely with a score of 95.3%, demonstrating near-optimal performance. Regarding mAP@0.5:0.95, YOLO v8x-OBB reached 70.9%, the highest value, while YOLO v8s-OBB followed with a score of 69.1%, reflecting its strong detection capabilities. Regarding computational complexity, YOLO v8s-OBB showed significant efficiency, with only 2.96 × 10¹⁰ FLOPs, much lower than the larger models like YOLO v8x-OBB, which required 2.64 × 10¹¹ FLOPs, indicating its efficiency in resource usage. Additionally, the model size of YOLO v8s-OBB was only 11.40 MB, considerably smaller than YOLO v8x-OBB's 69.50 MB, making YOLO v8s-OBB more suitable for deployment in resource-constrained environments. The influence of different network depths on the results was evident. Larger models, such as YOLO v8x-OBB, although slightly advantageous in some performance metrics, were constrained by their high computational complexity and larger model size. Therefore, considering the balance between performance and resource usage, YOLO v8s-OBB was the best base model for further improvements. It offerred high-performance detection while maintaining low computational complexity and model compactness.

In the ablation study, the detection performance of the YOLO v8s-OBB model under different configurations was evaluated to analyze the impact of each component on the overall model performance. The experimental results were shown in Table 3. The base model used only the YOLO v8s-OBB without the LSCD and CBAM modules. The results showed a precision of 94.9%, recall of 85.9%, and mean average precision (mAP@0.5) of 95.3%. The computational complexity of this configuration was 2.96 × 10¹⁰ FLOPs, and the model size was 11.4 MB, indicating that the base model already possessed high detection capability and efficiency. When the LSCD was introduced, the recall rate was increased significantly to 89.7%, although precision was decreased to 93.4%, and mean average precision dropped slightly to 95.0%. However, the computational complexity was decreased to 2.62 × 10¹⁰ FLOPs, and the model size was reduced to 9.5 MB. This demonstrated that the LSCD module effectively reduced the model's computational resource requirements while improving recall, making it more suitable for resource-constrained environments. Adding the CBAM to the YOLO v8s-OBB model results in a precision increase to 95.6%, a slight increase in recall to 86.0%, and a significant improvement in mean average precision to 96.4%. The computational complexity remained unchanged at 2.96 × 10¹⁰ FLOPs, and the model size stayed at 11.4 MB. This indicated that the CBAM module enhanced the model's ability to recognize key features, thereby improving detection accuracy. When both LSCD and CBAM were applied in the YOLO v8s-OBB model, the results showed the highest precision at 96.0%, an increase in recall to 90.1%, and a further improvement in mean average precision to 96.7%. The computational complexity was 2.62 × 10¹⁰ FLOPs, and the model size was 9.5 MB. The combination of these two modules maintained high precision while improving recall and effectively reduced computational complexity and model size, demonstrating that this combination achieved a good balance between enhancing model performance and resource efficiency. In summary, the ablation study showed that the CBAM and LSCD modules each contributed uniquely to improving the model's performance, and their combined use maximized detection performance, especially in resource-constrained environments, showing high application potential.

The experimental results on the test set were shown in Table 4, demonstrating that the improved YOLO v8s-OBB model had significantly enhanced detection performance compared with the original model. The improved model's precision was increased to 95.6%, compared with the original model's 92.5%, reducing false positives and allowing for more accurate identification of conidiophores and sporangium of cucumber downy mildew. The recall was also improved to 89.6%, up from 86.9% in the original model, indicating that the improved model can detect more targets and reduce the number of missed detections. For mean average precision (mAP@0.5), the improved model achieved 96.5%, higher than the original model's 94.9%, indicating high consistency and accuracy when detecting different targets. Under the more stringent mean average precision (mAP@0.5:0.95), the improved model scored 70.1%, while the original model achieved 69.1%, validating its robustness across different IoU thresholds, particularly in detecting targets within complex backgrounds and diversified scenarios. In summary, the improved YOLO v8s-OBB model outperformed the original model across key performance metrics, including precision, recall, and mean average precision. The enhanced ability to detect targets within complex backgrounds was particularly noteworthy. Furthermore, Fig. 6 presented a performance comparison of different models in detecting conidiophores and sporangia of cucumber downy mildew. In the S2ANet, H2RBox, and R2CNN models, severe cases of missed detections and repeated detections of conidiophores and sporangia can be observed, indicating limitations in precise localization. Compared with these models, the YOLO v5s-OBB and YOLO v7s-OBB models showed some improvement in bounding box accuracy, but they still struggled with identifying smaller or heavily overlapping targets. Some cases of missed detections for individual conidiophores and sporangia were also presented. In contrast, the improved YOLO v8s-OBB model outperformed all the displayed models, achieving the most accurate bounding box positioning and effectively distinguishing multiple closely located targets. Each target's bounding box was clear and precisely positioned, demonstrating high accuracy with minimal overlap. This performance proved that through refined feature extraction and optimized bounding box processing mechanisms, the practical utility and efficiency of the YOLO v8s-OBB model were significantly enhanced, making it more competitive and higher-performing in multi-target detection tasks within complex backgrounds.

Additionally, Fig. 7 showed the heatmaps generated by the original YOLO v8s-OBB model and the improved YOLO v8s-OBB model for detecting the sporangia and conidiophores of cucumber downy mildew. In the heatmap of the improved YOLO v8s-OBB model, the warmer tones (yellow to red) highlighted the areas that the model considered most likely to contain the target objects. With the introduction of the CBAM attention mechanism and the LSCD, the improved model focused more on these areas of interest, displaying more concentrated hot zones. The colors around each target were more focused and well-defined, demonstrating the model's excellent performance in accurately identifying and analyzing specific structures in microscopic images.

Quantitative detection system was proposed based on the improved YOLO v8s-OBB algorithm for identifying and counting the conidiophores and sporangium of cucumber downy mildew (Fig. 8). The system leveraged the efficient detection capabilities of the deep learning model to perform real-time detection of targets in static images and dynamic videos. The system firstly loaded the best weight file obtained from training the improved YOLO v8s-OBB model and then performed inference on the input images or videos. The model can effectively filter and accurately identify target objects by adjusting parameters such as confidence (conf) and intersection over union (IoU). In addition to drawing detection boxes and annotating labels, the model can also count the number of target categories based on the detection results and display the instance count for each category on the image through a counter, thereby providing an effective tool for the quantitative analysis of cucumber downy mildew. The system supported processing static images and videos, and it can obtain real-time data streams through a camera, enabling real-time updates of detection results and making it suitable for online monitoring in actual production environments. The user interaction interface was built on the Gradio framework, allowing users to upload images and video files or conduct real-time detection through a camera and display the detection results. Users can adjust detection parameters as needed to meet the requirements of different environments. This system offerred a convenient and efficient solution for disease detection in the agricultural field, helping growers to monitor crop health in real-time, take timely control measures, and improve agricultural production efficiency.