Detection And Analysis Of Citrus Maturity Based On Multi-Feature Fusion

Citrus occupies an important position in the Chinese market.In its storage and transportation work,the maturity of the fruit at the time of picking is an important factor affecting the quality of citrus fruit harvesting and the preservation effect.Effective detection of it is an important basis for grasping the timing of fruit picking and transportation to the fresh food market,in order to achieve accurate harvest and maximum profit.At present,most methods for determining maturity still rely on the experience of fruit growers,and individual differences in experience can directly affect the quality of citrus picking,which cannot guarantee the quality of citrus entering the market.Another commonly used detection method requires the destruction of the internal organization of fruits for internal quality detection technology.Although the accuracy is improved,it will cause certain consumption and losses,and cannot achieve large-scale inspection.In response to the above issues,this article proposes a fast and non-destructive deep learning method to determine whether citrus has reached the optimal taste and high-quality maturity level.This article takes citrus data from the School of Technology of Huazhong Agricultural University as the sample,and records citrus images and light intensity values as the research object.The color index is used as an indicator to measure the maturity of citrus for detection.Firstly,statistical features are extracted from citrus images,and the distribution differences between image features and lighting features during the maturation process are analyzed from a statistical perspective.Based on this,features with significant differences in stages are selected for inclusion in the indicator system.Then,a neural network model is constructed to combine image statistical features with lighting features for predicting color index.By comparing the effects of several classic regression models,two main types of models were selected: image based multimodal models and image statistical feature based multi feature fusion models,and the best performing model was determined.Based on this,the main work of this article is as follows:(1)Firstly,preprocess the data,align the data,segment the image,and normalize the lighting features.Then,based on previous research experience,10 statistical features were extracted from the four color spaces of the image: RGB,HSI,L * a * b *,and grayscale space.Combining the two features of light intensity: light intensity(negative)and light intensity(positive),a total of 12 features were analyzed to determine the statistical significance of the data.Subsequently,the ripening process of citrus is divided into three stages based on its color conversion rate: the green and astringent stage,the color conversion stage,and the picking stage.Observing the distribution and trend of 12 features from descriptive statistical analysis results,combined with non parametric tests,evaluate the significance of differences in features at different stages.The experimental results indicate that there are significant differences in 10 color space statistical features and 2 lighting intensity features at different stages,so an indicator system is constructed using these 12 features.(2)This article first uses complete images as input and compares VGG16,Res Net18,and Mobile Net＿The multimodal deep learning model and the single modal(image based only)model of V2 are implemented using Res Net18 and Mobile Net,respectively＿The efficiency of constructing a mid range fusion model in V2 is slightly better than that of a single modal model.However,during the training process,it was found that it was difficult to balance the learning rate and network complexity of the two modalities,making it difficult to improve the effectiveness of the fusion model.Thus,we try to extract the statistical features of the image color space,simplify the network structure,and reduce the risk of overfitting.(3)When constructing a multi feature fusion model,the first step is to compare the multi feature front-end fusion models constructed using image statistical features: the machine learning model includes K-nearest neighbors,multiple linear regression,naive Bayes,and XGBoost,while the deep learning model is composed of a three-layer neural network.Both machine learning and deep learning models outperform only using image statistical features when using fused features,indicating that utilizing information from both lighting and image statistical features can improve the model’s expression ability.When designing the middle end model and back end model network,considering that increasing the network complexity will increase the risk of overfitting,we use the advantages of convolutional neural network to design partition images to increase the number of features.In addition,partition images can also retain certain spatial information.Therefore,this article studies the performance of four uniform image segmentation methods in the model: 11 is the global image,22 is the equal uniform segmentation into 4 regions,44 is the equal uniform segmentation into 16 regions,and1010 is the equal uniform segmentation into 100 regions.By comparing the evaluation indicators R2 and MSE of the model,it was found that the backend mean fusion model of44 partitioned images and the mid-range fusion model of 44 partitioned images performed better on the training and testing sets than the other three methods of image partitioning.Among them,the backend mean fusion model of 44 partitioned images performed the best on both the training and testing sets.This indicates that dividing image regions is beneficial for the model to learn image features.Overall,this article selects the color index as a maturity measurement indicator,which can effectively reduce errors caused by inaccurate maturity grading.The multi feature fusion model constructed in this article has good detection ability,which can effectively predict the color index of citrus through image and lighting values.It can provide an effective method for predicting the color index of citrus and can be applied to non-destructive testing of citrus maturity,replacing time-consuming and labor-intensive manual testing and lossy detection methods.