| Faba bean is the third most important winter edible bean crop in the world,with high nutritional value.However,during the transportation,storage and processing of faba bean,the color of its shell will deepen,even rot and deteriorate,and the quality of bean grains will also be affected.In the early years,the artificial method was mainly used to classify and screen faba beans,which had the problems of high classification error rate and low efficiency.In order to improve the accuracy and efficiency of faba bean classification,this thesis uses deep learning algorithm to solve the defect classification task of Ezhou faba bean No.1,improves the classical models such as DenseNet,VGG and Goog LeNet,and proposes three faba bean defect classification algorithms.The accuracy of the model based on MSG-DenseNet-Cos Ann reached 98.43%;the accuracy of the model based on BRI-VGG-Cos Ann reached 98.71%.The accuracy of the model based on Mi At-Goog LeNet-Cos Ann reached 98.93%.Compared with the original algorithm model,the accuracy of the three improved algorithms is increased by 1.5%,1.85%and 2.36% respectively.The main work of this thesis is as follows:1.The feature extraction layer of dense convolutional neural network DenseNet121 is improved,and a multi-scale attention dense connection convolutional neural network model is proposed.The feature extraction part of DenseNet121 is composed of dense block Dense blocks.Because of its fixed 3 × 3 convolution kernel,the feature information under different receptive fields will be lost.In this thesis,the Inception multi-scale structure is introduced,and the MS dual-branch structure is formed with the Dense Block.The GAM module is embedded in the MS structure to enhance the extraction of feature information,and the MSG multi-scale attention dense connection block is designed.The simulation results show that the multi-scale densely connected attention module effectively enhances the feature extraction ability of the model.2.The feature extraction layer of convolutional neural network VGG is improved,and a multi-scale residual convolutional neural network model is proposed.VGG is a deep convolutional neural network.After multi-layer convolution,the feature dimension is getting smaller and smaller,which will lose some feature information and limit the accuracy of classification.Therefore,ResNet-Inception structure is introduced to replace some convolution layers in VGG,and the risk of feature loss is reduced by multi-scale residual structure.Add average pooling,Dropout,batch normalization,etc.,remove some fully connected layers,and reduce the trainable parameters of the network.The simulation results show that compared with the original model of VGG series,the BRI-VGG multi-scale residual model proposed in this thesis has obvious advantages,and the computational complexity of the network has also been significantly improved.3.ResNet18,ResNet34,Goog LeNet,DenseNet121 and other network models are improved.AttBn attention structure is introduced to fuse shallow features with deep abstract features,and four attention fusion models are designed.The Mi At-Goog LeNet model was designed by simplifying the best performing AttBn-Goog LeNet model in the fusion model.Finally,the cosine annealing algorithm with restart effect is introduced on MSG-DenseNet,BRI-VGG and Mi At-Goog LeNet to adjust the learning rate.The simulation results show that the model designed in this thesis has higher classification performance,and the training parameters of the model are greatly reduced. |
