| Image segmentation method is mainly divided into two parts.The first part is the traditional visual image segmentation algorithm,and the second part is the deep learning algorithm.Traditional image segmentation methods include gray threshold segmentation,edge-based segmentation and specific theory-based segmentation.Because of its low efficiency,complex algorithm and large space to improve the effect,it is seldom used at present.For the deep learning algorithm,the support vector machine(SVM)algorithm is commonly used as a classification algorithm,combining a variety of different algorithms,such as genetic algorithm(GA)or watershed algorithm,in order to reduce the complex kernel function and parameter selection problems caused by the simple use of support vector machine.It is more effective than using support vector machine model alone.However,there is still much room for improvement in image adaptability and recognition and segmentation of overlapping objects.And still cannot meet the actual use of efficiency and accuracy requirements.This algorithm mainly improves the traditional algorithm and develops a new practical application algorithm in the actual picking process of Lingwu Long Jujube.This paper uses and improves convolution neural network algorithm to detect individual Lingwu long jujube growth image,classifies the detected Lingwu long jujube,and achieves pixel level segmentation for each Lingwu long jujube in the image.For the first time,the improved neural network framework was applied to Lingwu Long Jujube image segmentation.In order to make Lingwu Long Jujube recognize in convolution neural network,the original convolution neural network framework algorithm is improved.The target detection framework of Lingwu Long Jujube based on top-level error optimization parameters is an improvement of the traditional convolution neural network algorithm in feature extraction framework and image segmentation framework structure,in order to adapt to the actual growth environment of Lingwu Long Jujube exclusively.The proposed method is superior to the traditional method in target detection and classification.In this paper,the traditional convolutional neural network algorithm is used to optimize the test pattern invocation network,feature graph generation network and target segmentation network as follows:(1)Neural network image segmentation algorithm is applied to Lingwu long jujube recognition,which is a valuable practical application topic.Achieved the expected segmentation effect test picture,the network accuracy and time performance are excellent basically maintained at the accuracy rate of 99%.Moreover,it can accurately segment the jujube area and jujube overlap in each picture,and the leaves are occluded under the influence of these complex natural environmental factors.(2)The test mode call network is rewritten as a whole.On the premise of satisfying the function of the original calling program,the GPU is used to realize the cyclic processing of more than three pictures under a single calling program,and the subsequent processing speed is increased by three seconds.And the robustness of the network framework is not affected.(3)The traditional convolution neural network feature map generation module has been improved and a symmetric fully connected convolution network has been proposed.The idea of symmetry and the idea of gradual deepening of feature depth extraction through small networks have been applied to feature map generation.According to the experimental viewpoint that Dropout can improve the effect of normalization and prevent over-fitting,the normalized data distribution and Dropout combination are applied to the residual network module for the first time.After the network modification is extracted from the above feature map,the improved symmetric full-connection convolution network improves both in stability,output rate and image preservation in non-interesting regions.In reasoning mode,the accuracy of the algorithm is as high as 95%,which is much higher than any traditional image segmentation algorithm. |
