Research On Fine-grained Image Recognition Algorithm Of Apple Leaf Diseases Under Complex Background Based On Deep Learning

Crop disease identification technology is of great significance to improve crop output efficiency and enhance crop output quality.Traditional disease identification methods rely on personal experience judgment,with slow identification speed and high misjudgment rate.With the development of computer technology,crop disease recognition methods based on traditional vision and deep learning have made some progress.However,the traditional crop disease image recognition methods need to extract the disease image features manually.The workload is large,and the extracted features can not represent the optimal features of the disease.At present,deep learning technology has an outstanding application prospect in the field of image recognition.As an important branch of deep learning technology,convolutional neural network can independently learn the feature information in the image without manual processing,and has good recognition accuracy and recognition speed.However,most of the existing studies focus on crop disease image recognition under simple background and crop disease coarse-grained image classification methods,which can not achieve high recognition accuracy in natural environment.In order to solve the problem of low disease recognition accuracy caused by large background noise and high fine-grained disease characteristics in actual field operation,this paper takes the fine-grained image of apple leaf disease under complex background as the research object,researches the classification algorithm based on the improved convolution neural network and convolution neural network fusion multi head self-attention mechanism,and realizes the high-precision fine-grained image classification of apple leaf disease.The main research contents are as follows:(1)Taking the apple leaf disease image as the research object,the fine-grained image classification algorithm of apple leaf disease under complex background is studied.Aiming at the problems of weak fine-grained feature extraction ability and poor ability to distinguish between background noise and leaf information of the existing network,through experimental comparison,the residual network(Res Net)is selected as the basic network to improve the information flow and fast projection method in the residual bottleneck block,and the pyramid convolution embedded in the hole is used to extract the multi-scale features of the disease image,so as to improve the classification accuracy of the model.The results show that the improved Res Net proposed in this paper can reduce the feature loss of the network,enhance the learning ability and feature extraction ability of the model,and the classification accuracy on the test set reaches94.99%.Compared with the classical convolutional neural network,the accuracy of improved Res Net on the original test set is improved by 3.52%-7.85%,and that on the expanded test set is improved by 2.12%-3.24%.It has stronger learning ability,faster convergence speed and higher recognition accuracy.(2)Aiming at the problem that it is difficult to extract global information by convolution operation,resulting in the weak ability of existing networks to recognize edge lesions,a fusion model of convolution neural network embedded in visual converter(Res Net-Vi T)is proposed in this paper.Experiments are designed to explore the effects of the number of layers of convolution neural network,the depth of visual converter and the number of heads of self-attention mechanism on the performance of the model.The local information of disease spots extracted by convolution and the global information of disease spots extracted by multi head self-attention mechanism are fused to realize the accurate identification of edge diseases.Finally,this paper integrates the improved Res Net model and Res Net-Vi T model in the research content(1),proposes im Res Net-Vi T model,and compares it with the improved Res Net model.The results show that im Res Net-Vi T has high classification accuracy and confidence.The average accuracy,accuracy,recall and F1 score on the test set are 95.44%,95.85%,95.44% and 95.64% respectively.In conclusion,the improved convolution neural network model imResNet-ViT model proposed in this paper can accurately recognize the fine-grained image of apple leaf disease under complex background.The results show that the model can provide theoretical support for apple leaf disease recognition and reference for the development of crop image recognition and detection instrument.It has certain theoretical research significance and application value.