In order to improve the accuracy and efficiency of intelligent identification of watermelon seedlings in different periods, an intelligent recognition method with improved Feature Pyramid Network (FPN) is proposed by using deep learning technology. Firstly, a network model was designed by combining the feature pyramid network model and the Res2Net model, and the Efficient Channel Attention (ECA) mechanism was used to assign different weights to spatial features, and the channel parameter sharing was adopted to reduce the computational complexity of the model. Then, the residual structure was used to optimize and improve the model, which solved the problem of network degradation that occurred when the network depth continued to increase without increasing training parameters. Finally, the deep separable convolutions were used in the fully connected layer to replace traditional convolutions, significantly reducing computational complexity and achieving a lightweight design. The experimental results on watermelon seedling leaves at different growth stages show that the proposed recognition method has higher recognition accuracy and the shortest computational time, compared with several more advanced recognition algorithms, the recognition rate is up to 96.84%, an equal error rate is only 0.54% and the mAP is up to 91.68%, and the computational time is as low as 112 ms, providing technical support for promoting the development of intelligent agriculture and achieving intelligent agricultural management decision-making.

为提高对不同时期西瓜幼苗智能识别的准确度和运行效率，采用深度学习技术提出改进特征金字塔模型（FPN）的智能识别方法。首先结合特征金字塔网络模型和Res2Net模型设计网络模型，利用有效通道注意力机制（ECA)赋予空间特征不同权重，采用通道参数共享的方式，降低模型的计算复杂度；然后采用残差结构对模型进行优化改进，在不增加训练参数的情况下，解决网络深度不断提升时出现的网络退化问题；最后在全连接层使用深度可分离卷积替换传统卷积，从而大幅减少计算量，实现轻量化的设计。对不同生长期西瓜幼苗叶片进行试验。结果表明：与几种较为先进的识别算法相比，提出的识别方法具有更高的识别准确度和最短的运算耗时，识别率达到96.84%，等误率仅为0.54%，平均精度mAP达到91.68%，运算耗时低至112 ms，为推动智慧农业的发展和实现智能化的农业管理决策提供技术保障。