基于短波近红外高光谱和深度学习的籽棉地膜分选算法

Film Sorting Algorithm in Seed Cotton Based on Near-infrared Hyperspectral Image and Deep Learning

采用膜下滴灌棉花种植模式,在机械采摘过程中地膜易混入籽棉,对后续棉花加工影响极大。地膜无色透明且无荧光效应,常规方法很难识别。为了解决地膜的分选问题,提出一种基于短波近红外高光谱和深度学习的籽棉地膜分选算法。首先,针对高光谱图像中地膜与非地膜像素点光谱特征区分不明显的问题,利用堆叠自适应加权自编码器逐层提取与输出相关的低维非线性高阶特征;然后,将此高阶特征作为分类器的输入,采用粒子群优化的极限学习机实现初步分类;最后,对分类结果进行类型合并,运用形态学方法以及连通域分析,剔除误识别区域,得到优化后的地膜分类结果。经仿真试验及现场测试,算法对地膜识别率达到95.5%,地膜选出率达95%,满足实际生产需求。

As the main cotton-producing province in China,Xinjiang has widely applied film-covering technology.In the process of cotton mechanical picking,a large amount of film is also collected along with the seed cotton.If the film could not be thoroughly separated,it would be subsequently transformed into the ginned cotton together,which would reduce the quality of the textile.However,it is difficult to identify the film by using traditional methods,because the film is colorless and transparent without fluorescent effect.In order to detect the film covering the seed cotton,a novel algorithm was proposed based on shortwave near-infrared hyperspectral imaging and deep learning.Firstly,considering the advantage of multi-channel and model complexity,the variable-wise weighted autoencoder was developed to weight hyperspectral image channel and transform them into low-dimension feature.Comparing with selecting one or deleting some channels directly,VW AE was used to achieve information that was more useful and less influence on the negative feature.Then,several variable-wise weighted autoencoders were stacked layer by layer to form deep networks,and a two-layer neural network combined with the BP algorithm was used to update the deep network weights.Next,the high-level features from the deep network were set as the inputs of an extreme learning machine(ELM)whose parameters were determined by a particle swarm optimization method.Finally,the classification results of the ELM were merged into film and non-film two classes by morphology and connected domain technologies.Simulation experiments and a field test were carried out to evaluate the performance of the proposed algorithm.The results showed that the recognition rate of the presented algorithm was up to 95.5%and the separating rate of the film was 95%,which met the actual production requirements.