Research And Implementation Of Plug Seedling Classification Based On Attention Mechanism

The classification of plug seedlings has important guiding significance for the automatic transplanting and grading of crops.Aiming at the problems that the current plug seedling data set is small and the traditional plug seedling classification algorithm is not high in accuracy and the calculation process is complicated,the plug seedling data set was made,and the self-made data set was trained and classified based on the improved convolutional neural network.The plug tray seedling image detection and recognition software system realizes the automatic detection and classification of plug tray seedlings.The main research contents are as follows.(1)In view of the lack of plug seedling data set,collect and produce the plant＿seed seedling image data set,which contains seven different crop seedling plug images,and use Mobile Net V2,Google Net,Res Net101 network to classify and train the Plant＿seed data set,The classification effects of different networks on the Plant＿seed dataset are compared and analyzed to provide experimental and theoretical basis for subsequent work.(2)In view of the low classification accuracy of the convolutional neural network algorithm and the poor ability to extract plug image features,an improved attention mechanism residual network plug seedling classification algorithm model was proposed,and the plant＿seed data set was analyzed.Training,design different attention mechanism fusion methods for experimental comparison,the experimental results show that the network Res Net34+CBAM＿basic＿conv which simultaneously integrates the dual-channel attention mechanism module between the residual module and the convolution block(conv*＿x)achieves the highest accuracy.Excellent 93.8%,proving that the dual-channel attention module is superior to the single-channel attention module.Aiming at the problem of missing seedlings caused by plug image segmentation,a random rectangle occlusion algorithm was used to occlude the plug image,artificially expand the number of missing image samples,and increase the model’s ability to recognize images with missing parts.Experiments show that the error rate of model classification can be effectively reduced after random erasing of images.(3)Aiming at the problem that the improved network model proposed in Chapter three has low classification accuracy for some subclasses of the Plant＿seed dataset.An improved visual Transformer attention mechanism based plug seedling classification algorithm is proposed.The algorithm uses the attention mechanism residual module proposed in Chapter three to extract image features,and converts the image features into the input feature vector of the Transformer network.In the Plant＿seed data Experiments on the set show that the algorithm can effectively improve the classification accuracy of subclasses,and at the same time increase the overall classification accuracy to 98.57%.Greatly improve the classification ability of the model.(4)Aiming at the indispensable problem of automatic thinning software identification and detection software system in the application of mechanical automatic thinning,a plug seedling classification detection and identification software system was developed based on the trained convolutional neural network.The test results show that the plug seedling identification software system based on convolutional neural network can realize seedling classification.