Study On Status Judgement Of ResNet Tomato Grafted Seedings Based On Lightweight

Grafting is one of the important means of artificial plant propagation.Grafting can not only improve the survival ability of plants,but also effectively improve the yield of plants.The most important stage in grafting is the healing of the grafted seedlings.But at present,the healing status of grafted seedlings is mainly judged by artificial eyes,which not only leads to slow judgment and subjective influence,but also lacks an accurate criterion.Therefore,this paper proposed a convolutional neural network to judge the status of tomato grafted seedlings during healing.The main research contents of this paper are as follows:(1)In order to solve the problem that the size of target objects in different healing states of tomato grafted seedlings was inconsistent,the fusion of multi-scale characteristic information was proposed to improve the accuracy of judgment of healing states of grafted seedlings.Through multi-scale convolution nuclear fusion and multi-scale characteristic figure merge two aspects,multi-scale feature fusion is proved the accuracy of network promotion effect,at the same time for multi-scale network and increase in the number of the problems of convolution kernels,based on the depth of separable convolution proposed two lightweight residual block structure,and choose the better network performance through testing,Effectively improve the speed of the network.(2)In order to realize better application of network model in mobile devices,Channnle Split is introduced to lightweight Res Net network.Firstly,two kinds of lightweight structures are designed for Res Net residual block structure,and the one with better performance is selected.Then,the influence of different network widths and depths on the number of network parameters and network performance was further studied to balance the number of network parameters and network performance and select the network structure with the best performance in the tomato graft status judgment task.Finally,experiments verify that the classification accuracy of the lightweight network is higher and the number of network parameters is less,achieving the balance between the network accuracy and lightweight.(3)In order to further improve the classification accuracy of the network,there is a small difference between different categories in the healing state of tomato grafted seedlings,which is not conducive to the accurate judgment of the healing state of tomato grafted seedlings.Based on Res Net50 network is introduced into the SE attention mechanism module,at the same time,the global average for SE module pooling layer for feature extraction,there is no good to extract features of different scales,will be replaced by global average pooling layer pyramid pooling,pyramid after pooling,introduced by experimental verification,can effectively improve the classification performance of the network;At the same time,in view of the introduction of pyramid pooling will increase the number of network parameters,the lightweight SE module based on Channle Split is proposed.Through experimental verification,the improved attention mechanism module not only reduces the increase of the number of original basic network parameters,but also improves the classification performance of the network.