Fish Freshness Detection Design Based On Deep Learning

With the arrival of a well-off society,people’s living standards have been improved.The awareness of food safety and quality has also been increasing,making it a focus of concern for consumers.Fish,as a highly perishable product,is difficult to preserve even in a refrigerated environment.The decay of fish can cause economic and food health problems,so fish freshness testing is an effective means of ensuring fish quality.Currently,traditional freshness testing methods are complex,have low recognition rates,and require expertise,making it impossible to meet practical needs.Therefore,a fast,simple,and accurate freshness testing method is of great significance.This thesis proposes the following research results through a large amount of data collection and experimental verification:1.In the study of fish species identification,this thesis collected 4681 images of 10 fish species.By training and optimizing four deep convolutional neural network models,AlexNet,VGG-19,ResNet-18,and ResNet-50,it was found that the ResNet-50 model has better fish image classification ability,with a verification accuracy of 97.42% and an average accuracy of 96.57%.At the same time,this thesis also verified transfer learning and data augmentation algorithms,and the results showed that the accuracy of the model was significantly improved after transfer learning and data augmentation.2.This thesis also collected and studied fresh yellow croaker image data and obtained a total of 1351 images of fresh yellow croaker freshness.The volatile basic nitrogen content(TVB-N)in the body of fresh yellow croaker was tested experimentally at 4°C,and it was found that the shelf life of fresh yellow croaker at 4°C was 5 days,with a shelf life of 2 days for first-class products.At the same time,this thesis proposed the change of TVB-N in the yellow croaker body under 4°C within 7 days.The TVB-N values of the seven fish body detections were 10.06mg/100 g,12.36mg/100 g,15.26mg/100 g,21.76mg/100 g,29.46mg/100 g,47.36mg/100 g,and 71.83mg/100 g.3.This thesis obtained a fish freshness recognition method based on ResNet-50 through training the fresh yellow croaker image dataset,with an accuracy of 97.84% in the validation set.To better evaluate the effectiveness of the model classification accuracy,the F1-score parameter was introduced,and the results showed that the F1-score parameter of ResNet-50 in the recognition of freshness on various levels was higher than that of other networks.4.This thesis made changes to the above optimal model ResNet-50.Firstly,three small convolution kernels were used instead of one large convolution kernel in the Conv1 layer.Secondly,some convolution downsampling was replaced by average pooling.The results showed that under the condition of ensuring that the parameter quantity is basically unchanged,the validation accuracy of the ResNet-50＿1 model reached 98.46%,the test accuracy reached 96.19%,and the test speed was 0.018s/image,all higher than the original ResNet-50 network.5.Based on the above,a fish image freshness detection system was designed,which is based on the PyQt5 framework design of a GUI visualization interface.The system can perform fish species identification,known fish freshness detection,and unknown fish freshness detection on the images to be tested.The research results of this thesis demonstrate that fish species identification and freshness detection methods based on deep learning can conduct fish freshness testing quickly,simply,and accurately,and have significant practical value.