In order to improve the accuracy of weed recognition in cane fields under natural environment, a weed detection algorithm based on improved Faster Regionbased Convolutional Neural Network (Faster RCNN) was proposed. Firstly, in the feature extraction stage, the balanced feature pyramid module was used to balance the semantic features at all levels to strengthen the extraction of deep features of weed images. Secondly, the dynamic label assignment was used to dynamically adjust the label prediction threshold of the network to solve the problem of scarcity of positive samples in the early stage of training. Finally, soft nonmaximum suppression was used to optimize the model, which was able to reduce the missed detection of singletype targets and improve the positioning accuracy of targets by improving the nonmaximum suppression of the original model.The experimental results showed that the mean average precision of the optimized algorithm reached 81.3%, which compared with the original Faster RCNN algorithm, the precision was improved by 6.2%, and the average test time for each image was 0.132 s. There were 6.5%, 4.7%, and 5.1% improvements in the average precision of the intersection over union of 0.5 and the across scale of small and medium, respectively. The proposed algorithm has high detection precision and stability, which can meet the needs of sugarcane field weed detection in complex natural environment.

为提高自然环境下蔗田杂草检测准确率，提出一种基于改进的Faster RCNN的蔗田杂草检测算法。在特征提取阶段使用BFP模块均衡各级语义特征来加强对杂草图像深层特征的提取；采用DLA策略动态调整网络的标签预测阈值，解决训练前期正样本稀缺问题；使用SoftNMS对模型进行优化，通过改进原模型的NMS减少单类目标漏检并提高目标定位精度。试验结果表明，优化后算法的mAP值达81.3%，与原Faster RCNN算法相比，精度提升6.2%，平均每幅图像测试耗时0.132 s，且在AP50、APs、APl指标上分别有6.5%、4.7%、5.1%的提高。改进后的算法具有较高的检测精度和稳定性，可以满足复杂自然环境下的蔗田杂草检测需求。