Research On Image Segmentation Of Multispecies Fruit Flower Based On Deep Learning

In the process of fruit production,the statistics of flower number is often the key factor for the later fruit management decision-making.Extracting flower contour from flower image and estimating the number of flowers will be beneficial for fruit growers to cultivate fruit trees at flowering stage.Although the existing flower segmentation methods can segment the target flower,its accuracy and practicability need to be further improved,and the range of use is small,it is only an effective manual image processing technology under specific conditions.In most methods,in order to adapt to the variety of flowers and fruits,weather conditions,camera position(distance and angle)relative to the orchard,the color and size threshold parameters of these algorithms must be readjusted.Besides,it can only separate one kind of flower,and can't adapt to many kinds of flowers.There are many objects in flower image and the problem of color similarity between flower and background,which makes a recognition process to construct a complete segmentation process before segmentation.This paper proposes an effective and automatic segmentation method for apple flower,peach flower and pear flower high-resolution images with different backgrounds,which mainly uses the fine-tuning CNN network to identify the flower region,and uses the shape constrained level set model to fine-grained segmentation.The main work is as follows:(1)Aiming at the problems of complex background,light change and variety of flower image,this paper proposes a new method of flower image segmentation,which has the following three characteristics: first,it is an automatic system;second,it is robust to the change of complex environment and light;third,it can be extended to many kinds of fruit flower image segmentation,such as peach blossom,pear blossom.In the training network,we only use a single apple flower image for training,which can be applied to other kinds of flowers without data preprocessing or specific multiple training.Then,a shape constrained level set model is proposed as the post-processing module of the network to refine the rough segmentation results of the network output.After finite iterations of the evolution equation,we can get the accurate contour of the target flowers.(2)Because the memory space needed for full convolution network calculation increases exponentially according to image resolution.Therefore,in the segmentation method,we cut the high-resolution image into a specific small image,and use the partial overlapping image cutting method to reduce the impact of the artificial cuttingboundary.(3)In this paper,the segmentation method is simulated and compared with three typical segmentation methods qualitatively and quantitatively.The results show that under the conditions of different light,different environment and different varieties,the visual effect of our method is the best,and the segmentation accuracy is the highest.The accuracy of four different data sets is as low as 82% and as high as 94%.(4)In order to estimate the number of flowers by segmentation results,we fit the relationship between the flower coverage and the number of flowers,and find their linear correlation,so we can predict the number of flowers in the image by the flower coverage.Our method only needs a high-resolution image of fruit trees to estimate the number of flowers in one minute,while it may take an average of 50 minutes to calculate the number of flowers per tree.