Activity Identification Of Crucian Carp Based On Image And Sound

The activity of freshwater fish reflects its health and behavior.Monitoring and identifying the activity of freshwater fish is helpful to detect the abnormal behavior in the aquaculture,so as to give early warning and take remedial measures in time,reduce the loss in the aquaculture,and improve the yield and efficiency of aquaculture.At present,the monitoring of freshwater fish activity is mainly through computer vision,which is easy to be affected by light,water quality,aquaculture density and other factors,and the scope of application has certain limitations.In view of the above problems,this paper proposed a passive underwater acoustic method to identify the activity of crucian carp.The voice and video signals of crucian carp were collected,and the activity level of crucian carp was classified by video signals.At the same time,the features of voice signal were extracted,and the activity recognition models of crucian carp were established.The main research work and conclusions were as follows:(1)The classification method of crucian carp activity was established.The single-stage target detection network in deep learning was used to detect the target of crucian carp,and multiple indexes were selected to evaluate the results.The F value and average accuracy were 0.58 and 0.87 respectively,which showed that the method had high precision for the detection of crucian carp target.The swimming trajectory of crucian carp was extracted,and then the swimming speed was calculated.The k-means clustering method was used to classify the activity level.The p value of student's t test was all less than 0.01,which indicated that the three activity levels were significantly different,and the activity level classification method adopted was more reasonable.(2)The sound signal traits of crucian carp were extracted and analyzed.The collected sound signals were filtered by Wiener filter,and the characteristics of sound signals were extracted,and the significance analysis of the acoustic signal characteristics of crucian carp with different activity levels was carried out.The optimal number of wavelet packet decomposition layers was determined as six,and 54 features with significant difference were selected,such as short-term average energy,45 six level wavelet packet decomposition frequency band energy,six average Mel frequency cepstrum coefficients,main peak frequency and main peak value of power spectrum.The first 27 principal component features with cumulative contribution rate of 99% were selected by correlation analysis and principal component analysis.(3)The identification models of crucian carp activity based on passive underwater acoustic signal were established.The 54 features and the first 27 principal component features were used for modeling,Fisher linear discrimination,support vector machine and random forest classification model were established,and the model was evaluated to compare the recognition effect of different models.The optimal activity recognition model of crucian carp was the random forest classifier model,and its classification accuracy and Kappa were 98.61% and 0.98 respectively.The experimental result showed that there was correlation between the acoustic signal and the activity of fish,so the passive underwater acoustic signal could be used to identify the activity of crucian carp.This study was of great significance for the in-depth analysis of the relationship between the acoustic signal and behavior of freshwater fish,and the behavior monitoring in the aquaculture.