Research On Wild Mushroom Identification Method Based On Deep Learning

With the rapid development of Yunnan’s tourism industry,wild mushrooms have become a must-try local delicacy for tourists.However,most people’s ability to identify wild mushroom species is limited.As a result,poisoning incidents caused by the ingestion of toxic wild mushrooms are common.To achieve accurate identification of wild mushrooms,existing research uses deep learning technology for wild mushroom recognition and classification.However,the quality of the datasets is highly variable,which affects the classification results.In addition,the traditional convolutional neural network has some shortcomings in feature extraction,such as the insufficient ability to extract local features of objects and the lack of lowlevel features in deep networks,which lead to unsatisfactory classification effects.Meanwhile,some research attempts to apply weakly supervised fine-grained image classification algorithms to solve the wild mushroom recognition and classification.However,this kind of algorithm is easy to interfere with background and irrelevant regions in some scenes,which leads to the inaccurate location of the target subject,thus affecting the classification effect.In order to improve the effect of fine-grained image classification and achieve efficient and accurate identification of wild mushrooms,this thesis carries out research from the following aspects.(1)In order to solve the problem that traditional convolutional neural networks lack the ability to extract local features of objects,and the lack of low-level features in the deep network leads to the unsatisfactory classification effect,this thesis proposes an MSFRes Net network model that integrates multi-scale feature information.In this model,a multi-scale attention mechanism is used to enhance the ability of the network to extract local fine features of objects,and the shallow feature map rich in low-level texture details is used as the input of the deep network through short connections,to guide the model to learn the underlying features and improve the recognition effect of the model.Experimental results on the public fine-grained image dataset FGVC-Aircraft show that the proposed algorithm effectively improves the recognition performance of fine-grained images.Moreover,for the wild mushroom recognition problem,MSFRes Net achieves a classification accuracy of 99.13% on the validation set,an improvement of 0.47% over the Res Ne Xt50 model.Heatmap results also indicate that the proposed model significantly reduces the interference of background information on the recognition of the main object,allowing the network to focus more on the primary features of the object,thus improving the recognition accuracy.(2)In order to solve the problems of the existing fine-grained image classification algorithms based on location recognition,such as the weak ability of basic network identification and positioning,the large computation amount of the overall network,and the poor usability of the model,this thesis proposes a two-branch network,D-Res MSF-Net,which integrates multi-scale feature information.By introducing a multi-scale attention mechanism and a part-level image generation module,the algorithm improves the network’s ability to locate and recognize key object regions and strengthens the extraction of local features.D-Res MSFNet learns both global and local feature parameters of objects from the original and part-level images,thereby improving the performance of the network.Experimental results on publicly available fine-grained image datasets show that the proposed algorithm can significantly reduce the amount of computation and improve network usability and training speed while ensuring accuracy.At the same time,the model visualization results also show that the D-Res MSF-Net model can accurately identify the key local regions in the localization image,and improve the ability of the network to extract the local features of the image.In the identification and classification of wild mushrooms,a classification accuracy of 99.60% was achieved on the verification set.(3)Based on the proposed model in this thesis,we have designed and implemented a wild mushroom recognition and detection system.The system is developed using the Python programming language and is built upon the proposed D-Res MSF-Net model to implement its core functionality.This wild mushroom recognition and detection system provides a convenient visual detection method that can accurately detect and identify wild mushroom images and provides users with an efficient and practical solution for wild mushroom identification.