Research On Aquatic Freshness Detection Method Based On Data Mining

With the growth of people’s demand for aquatic product quality,enterprises pay more and more attention to the freshness change of aquatic products in the transportation process,and need to use freshness detection,which plays a vital role in the production,transportation and sales of aquatic products.In the freshness detection of aquatic products,the commonly used method is volatile salt-based nitrogen detection,but this method will contaminate the sample,and the number of tests is small,and it cannot be fully covered.Based on machine vision and deep learning,this dissertation identifies and classifies the freshness of fish through fish eye information.This dissertation focuses on the embedded implementation of freshness detection and freshness detection using deep learning.The following is the research work of this dissertation :1.Machine vision is combined with deep learning and attention mechanism and applied to the task of aquatic freshness detection and classification.First,find the appropriate data set and the pixel factor parameters suitable for this experiment,and perform data set enhancement operations to prevent overfitting.Then,by comparing the three network model experiments,Mobilenet V1 is selected as the basic network.Secondly,the performance parameters of the Mobilenet V1 network with alpha as the default value and the model with the Self-Attention module separately added at the end of the Mobilenet V1 network are compared to verify the feasibility of the new model.Then,the alpha values of all models are reduced,and the performance comparison is performed to obtain training files with smaller files and better performance.Finally,the third-party platform is used to complete the training of the data set,and the model is run in the cloud to facilitate subsequent comparative experiments.2.Integrating deep learning models with embedded devices.First,modify the convolution kernel and part of the activation function in the model,and then prune the trained h5 file to obtain a file that occupies less memory.Secondly,evaluate and compare the model file obtained after pruning with the model file before pruning to determine the feasibility of the pruning model.Then,convert the h5 format file into the tflite format file,and then compile the tflite file into the kmodel file with the NCC tool,and burn the final file to the K210(Canaan Technology edge computing chip)development version,and write a python script file for model performance testing.Finally,a control group experiment was designed to compare the results of each group and select the most effective model.Design related experiments,compare different classification methods,and conclude that the research method in this paper is more suitable for freshness detection.