Research On Identification Technology For Fruit And Vegetable Freshness Based On Deep Learning

Freshness recognition of fruit and vegetable images can not only help reduce the wasteful situation of fruit and vegetable picking,storage,transportation and sales,but also effectively reduce production costs and sales prices,providing better quality and more affordable fruit and vegetable products.In recent years,with the development of artificial intelligence technology,relevant algorithms based on deep learning have become a research hotspot.The study of deep learning-based fruit and vegetable freshness recognition technology can provide a way to reduce labor costs and standardize the implementation of standards in the food field.In this paper,we propose a deep learning-based fruit and vegetable freshness recognition method,which achieves accurate recognition of fruit and vegetable image freshness while completing the research of knowledge distillation and model quantification techniques.The main research contents are:(1)A fruit and vegetable freshness classification dataset with different sample sizes was constructed.The 10% and 25% datasets were extracted to obtain FruitVeg-A and FruitVeg-B class datasets respectively,and PCA(Principal Component Analysis)color enhancement was performed on the original dataset to obtain the FruitVeg-Aug dataset containing a total of180,168 samples of 33 classes,which completed the The overall color feature analysis of fruits and vegetables was completed,which provided the data basis for fruit and vegetable freshness recognition model optimization and algorithm design.(2)The parameters of the fruit and vegetable freshness classification model were optimized based on a small number of FruitVeg-A(5977 samples)and FruitVeg-B(14989samples)class datasets.Grid tuning of the two parameters,modulation factor and learning rate,using the Mobile Vi TV1 model was effectively improved before optimization on the FruitVeg-A2 and FruitVeg-B1 datasets.(3)Benchmark performance and time consumption analysis of 22 lightweight models such as Mobile Net,Shuffle Net and Efficient Net and 3 conventional models such as VGG16,Res Net50 and Dense Net121 were completed based on a large number of FruitVeg-Aug(180,168)datasets.Among the lightweight models,model Mobile Vi T-XS has the highest accuracy of 96.71%;among the conventional models,model B-Dense Net121 has the highest accuracy of 98.22%.(4)Efficient Net V1-B0,Mnas Net,Mobile Vi TV1-XS,Shuffle Net V1,and ESPNet V2 were selected as student models by comparing the number of model parameters and accuracy.By comparing the model accuracy,B-Res Net50 and B-Dense Net121 after migration learning were selected as the teacher models.The optimal distillation temperature was determined by parameter search using B-Dense Net121 and Mnas Net,and the accuracy of the student model was effectively improved using knowledge distillation.After transferring the Onnx inference model and Tensor RT quantization model,the inference time consuming of the final student model was effectively reduced.(5)In this paper,we used Cleanlab for data cleaning,used B-Res Net50 and B-Dense Net121 models to calculate the FruitVeg-Aug test set health,and filtered out the top42 most likely wrong labeled samples to complete the analysis of error-prone samples.This paper also completes the iterative process of the convolutional layer of the Mobile Vi T model for the training process and the t-SNE dimensionality reduction visualization of the teacher model to extract all features of the test set in FruitVeg-Aug.In this paper,model performance parameter optimization,knowledge distillation and quantization operations are performed for small and large quantity datasets,respectively.Using the lightweight model as the student model and the regular model migration parameters as the teacher model steadily improves the performance of the model.Moreover,this paper also takes into account the characteristics of uneven number of parameters and existence of hard-to-identify samples in small-volume datasets,and performs parameter search and loss function optimization.Finally,using the techniques of inference model optimization and quantization model commonly used in engineering projects,we effectively optimize the time-consuming and memory occupation problems of the model in actual deployment tests,and make a more complete exploration and summary of the optimization,training and deployment of large quantity datasets.