Research On Cotton Recognition Method Based On Adaptive Segmentation Algorithm And Transfer Learning

Cotton is one of the main cash crops in China.Although the mechanical cotton picking method is lower in cost and higher in efficiency than manual cotton picking,it also leads to the problem of low quality and low yield of cotton.Therefore,the research on intelligent cotton-picking robot is the inevitable trend of cotton picking in the future.Visual positioning system is the most important part of intelligent cotton picker.In this paper,the recognition and positioning algorithm of cottons in natural scene is studied.In order to locate the cotton target in the natural scene accurately and efficiently,the following problems need to overcome.In the segmentation result,in addition to a single cotton object,there are overlapping cottons,cottons covered by branches and leaves,and the background that is mistaken as cotton peaches.Therefore,it is necessary to classify the segmentation result and study the appropriate image classification algorithm.In order to return the actual number of overlapping cottons and its corresponding coordinate position,the overlapping cottons separation algorithm needs to study.This paper mainly focuses on the three problems faced by cotton recognition:(1)Analyze the color value distribution characteristics of cotton field images under natural scenes in RGB and HSV(Hue,Saturation,Value)modes,and combine the effects of some classical image segmentation algorithms in cotton field image segmentation,aim for the cotton field image characteristics and put forward cotton adaptive segmentation algorithm.Firstly,The cotton field image is divided into sunshine and shadow parts according to the S value distribution characteristics of HSV space.Secondly,the sunshine image is segmented by the improved Otsu algorithm,and the shadow image is segmented by the threshold segmentation method based on color statistics and the improved Otsu algorithm.Combining the binary graph of the segmentation result of the sunshine and shadow sides.Finally,the paper applied the morphological algorithm to optimize the result and realized the segmentation and recognition of cotton peaches in the natural scene.The statistical results of the segmentation algorithm in recognition rate,false recognition rate and segmentation time prove the advantages of the segmentation algorithm in recognition rate and speed.(2)The transfer learning model based on AlexNet,GoogLeNet and resnet-50 is used to classify cotton dataset,and the classification effect is compared with the effect of CNN(Convolutional Neural Networks)trained with cotton dataset.It is proved that the deep convolutional neural network trained with ImageNet dataset has the ability to extract abstract features of general objects,and has a good classification effect on small cotton dataset,while the CNN classification modle trained with cotton dataset directly will generate overfitting.At the same time,transfer learning method can avoid making large datasets and retraining a deep CNN model,which saves a lot of time cost and computing resources.(3)On the basis of verifying the ability of transfer learning model to extract the features of cotton dataset,a new algorithm based on transfer learning model feature extraction and classifier is proposed,which further improved the accuracy of the algorithm.By comparing the classification effect and time consumption of various methods,AlexNet is selected as the feature extraction model and ELM(extreme learning machine)as the classifier,so that the classification algorithm can meet the requirements of classification accuracy and efficiency at the same time.(4)In view of the color and morphological characteristics of overlapping cottons,the separation method of overlapping cottons is studied.