SFC_DeepLabv3+: A Lightweight Grape Image Segmentation Method Based on Content-Guided Attention Fusion

In recent years, fungal diseases affecting grape crops have attracted significant attention. Currently, the assessment of black rot severity mainly depends on the ratio of lesion area to leaf surface area. However, effectively and accurately segmenting leaf lesions presents considerable challenges. Existing grape leaf lesion segmentation models have several limitations, such as a large number of parameters, long training durations, and limited precision in extracting small lesions and boundary details. To address these issues, we propose an enhanced DeepLabv3+ model incorporating Strip Pooling, Content-Guided Fusion, and Convolutional Block Attention Module (SFC_DeepLabv3+), an enhanced lesion segmentation method based on DeepLabv3+. This approach uses the lightweight MobileNetv2 backbone to replace the original Xception, incorporates a lightweight convolutional block attention module, and introduces a content-guided feature fusion module to improve the detection accuracy of small lesions and blurred boundaries. Experimental results show that the enhanced model achieves a mean Intersection over Union (mIoU) of 90.98%, a mean Pixel Accuracy (mPA) of 94.33%, and a precision of 95.84%. This represents relative gains of 2.22%, 1.78%, and 0.89% respectively compared to the original model. Additionally, its complexity is significantly reduced without sacrificing performance,the parameter count is reduced to 6.27 M, a decrease of 88.5% compared to the original model, floating point of operations (GFLOPs) drops from 83.62 to 29.00 G, a reduction of 65.1%. Additionally, Frames Per Second (FPS) increases from 63.7 to 74.3 FPS, marking an improvement of 16.7%. Compared to other models, the improved architecture shows faster convergence and superior segmentation accuracy, making it highly suitable for applications in resource-constrained environments.