A novel method for counting wheat ears

As an effective method for monitoring and evaluating crop yield, wheat counting using image data plays a vital role in agriculture. However, current methods face significant challenges that hinder their effectiveness. Many existing models require high volumes of data, which can be difficult to obtain and manage, leading to inefficiencies in real-world applications. Additionally, these methods often struggle with inaccurate counting, primarily due to issues such as overlapping wheat ears, occlusion, and varying environmental conditions that complicate image analysis. To address these challenges, this paper presents a wheat counting method based on LC-DeepLab. By integrating the lightweight MobileNet V2 as the backbone, the approach enhances data processing speed and recognition accuracy. The model improves feature extraction through concatenated atrous convolutional layers in the pooled pyramid module. A localization-based counting loss function (LC) is also utilized to further optimize accuracy, ensuring more reliable identification of wheat ears even in complex scenarios. Experimental results on the Global Wheat dataset demonstrate the model's effectiveness, achieving a precision of 92.20%, recall of 92.56%, and mean average precision (mAP) of 90.89%. The proposed method outperforms benchmark models, increasing precision, recall, and mAP by 2.21, 1.74, and 1.93 percentage points, respectively, while meeting real-time detection needs. The LC-DeepLab model exhibits superior robustness in handling complex situations, such as occlusion and overlap, underscoring its potential application in intelligent agriculture.