| Cotton is one of the most important economic crops in the world,with large yield and low production cost.The price of cotton products is relatively low.Cotton pigment glands are the main storage organs of gossypol,which is widely used in agriculture,medicine and other fields.It is often used to control pests and rodents,as well as make drugs to inhibit cancer.However,gossypol has certain toxicity.Therefore,it is of great significance to predict the content of gossypol in cotton plants to improve the economic value of cotton byproducts.The number of pigment glands in cotton affects the content of gossypol in cotton plants.Statistical analysis of the number of pigment glands is an effective method for estimating the gossypol content in cotton plants.Cotton pigment glands are extremely small and densely distributed,making manual counting methods difficult and easily influenced by the experience of the counter.With the continuous improvement of computer technology,computer vision technology is more and more widely used in the field of agriculture.Automatic recognition of targets in images through machine learning provides new ideas for the recognition and counting of pigment glands.In order to realize the rapid and accurate counting of pigment glands in cotton leaves,this paper uses machine vision deep learning algorithm to segment and count the pigment glands in cotton leaves.The main research contents are as follows:(1)A set of cotton leaf image acquisition device was built to obtain RGB images of cotton leaves and establish sample datasets.At the same time,according to the density of pigment glands in cotton leaves,the collected cotton leaves were classified into three density levels:low,medium and high.(2)To address the issues of low resolution and difficult feature extraction of small targets,based on the classical U-Net model,an interpolation-pooling network structure IPP Net is proposed.The network has a good performance in the densely distributed pigment gland segmentation task.The accuracy rate is 0.967,and the evaluation indicators mIoU,Precision,Recall and F1-score are 81.81%,80.04%,80.04%and 80.04%,respectively.Because the structure adopts the idea of first enlarging the image and then reducing the image,the calculation amount of the model is greatly increased.To reduce the computational complexity of the model,further enhance the feature extraction ability of pigment glands,and reduce the computational complexity of the model,the AttU-Net model based on the UNet network was proposed.The recurrent residual module is used to replace the original convolutional layer,which increases the network depth and reduces the gradient vanishing.The ResPath connections are used to connect the layers corresponding to the encoding and decoding parts,and the DUpsampling method is used to reduce the loss of pigment gland features during down-sampling.Finally,adding an attention module before ResPath connection enables the network to locate target pixels more accurately.The experimental results show that the improved AttU-Net has an accuracy of 0.981,and the evaluation indicators mIoU,Precision,Recall,and Fl-score have reached 82.67%,84.40%,86.30%and 85.30%respectively,which are 10.37%13.87%,28.04%and 21.5%higher than the classical U-Net model.The results show that the AttU-Net model has high detection accuracy in the segmentation task of densely distributed pigment glands,and greatly reduces the number of parameters in the model.(3)The pigment glands in cotton leaves were counted,which provided the basis for predicting the content of gossypol.In order to verify the counting ability of the two improved algorithms for different density pigment glands,IPP Net and AttU-Net were used to count and analyze the pigment gland datasets of three different density levels,and the results were fitted with the manual counting results.The results showed that the determination coefficients R2 of the counting results of AttU-Net model and manual counting results in three different density pigment gland datasets reached more than 0.9,which was closer to the manual count results.The counting results of the test set were compared with those of U-Net and ImageJ counting software.The counting results of AttU-Net are more accurate and can replace manual counting to a certain extent.The semantic segmentation network proposed in this paper performs well in the segmentation and counting of cotton pigment glands in cotton leaves,and has certain guiding significance for evaluating the content of gossypol in different varieties of cotton leaves. |
