长绒棉棉铃成熟度仿生类脑分类方法研究

Research on bionic brain‑inspired classification method for maturity of long‑staple cotton bolls

快速准确识别长绒棉棉铃成熟的不同阶段,对长绒棉种植的智能化、装备化管理具有重要意义。针对现有方法在长绒棉棉铃成熟度分类辨识易受复杂棉田背景、阴影、强光和叶片遮挡等因素影响识别率低的问题,提出将模拟生物视觉皮层的信息处理机制与HMAX模型融合的类脑分类方法,实现大田环境下棉铃不同成熟度的快速、高效识别。首先,采用相机获取长绒棉不同成熟阶段的图像信息,以花铃期、裂铃期、吐絮期、停止生长期4个生长阶段为对象,构建中小样本的棉铃数据集;其次,模拟视网膜神经节细胞的信息处理机制以提升HMAX模型的检测速度与精度,提出基于改进HMAX模型的类脑识别算法;最后,为探究各模型在非清晰数据集上的表现,采用高斯模糊方法将测试集转换6次,以HMAX、HHMAX、SHMAX作为对比,评估改进HMAX模型性能。试验结果表明,在原始数据集下,改进HMAX模型的总体准确率为95.3%,相比于HMAX、HHMAX、SHMAX模型分别高出15.1、9.2和6个百分点;在错误分类中,由于吐絮期和停止生长期特征相似,造成错误识别的概率最大;在非清晰数据集下,HMAX、HHMAX以及SHMAX退化指数分别为8.21%、7.935%和11.21%,改进HMAX模型总体退化指数为5.92%。研究结果表明:在分类精度、模糊图像输入等方面,改进HMAX模型能够较好地满足实际生产中棉铃不同成熟阶段分类识别的实际需求。

Rapid and accurate identification of different stages of long‑staple cotton boll maturity is important for intelligent and equipped management of long‑staple cotton cultivation.In order to address the problem that the existing methods in long‑staple cotton boll maturity classification recognition are easily affected by complex cotton field backgrounds,shadow,bright light and leaf shading,this paper proposes a brain‑like classification method that integrates the information processing mechanism of simulated biological visual cortex with the HMAX model to achieve fast and efficient recognition of different boll maturity in a large field environment.Firstly,a camera is used to obtain the image information of different maturity stages of long‑staple cotton,and four growth stages,namely,the boll stage,boll splitting stage,linting stage and cessation stage,are used to build a small and medium‑sized boll dataset.Secondly,the information processing mechanism of retinal ganglion cells is simulated to improve the detection speed and accuracy of the HMAX model,and a brain‑like recognition algorithm based on the improved HMAX model is proposed.Finally,in order to explore the performance of each model on non‑clear data sets,the test set was transformed six times by using the Gaussian fuzzy method,and HMAX,HHMAX and SHMAX were used as comparisons to evaluate the performance of the improved HMAX model.The experimental results showed that the overall accuracy of the improved HMAX model on the original test set was 95.3%,which was 15.1,9.2 and 6 percentage points higher,respectively,compared with the HMAX,HHMAX and SHMAX models.In the misclassification,the probability of false identification was the highest due to the similarity of the characteristics of the situation and cessation growth periods.Under the non‑clear data set,the degradation indices of HMAX,HHMAX and SHMAX were 8.21%,7.935%,and 11.21%,respectively,and the overall degradation index of the improved HMAX model was 5.92%.The results show that the improved HMAX model can better meet the practical needs of classification and recognition of cotton boll at different maturity stages in actual production in terms of classification accuracy and fuzzy image input.