| With the increasing requirements for aquatic products,the recent decades witnessed the scale of aquaculture market expanding continually.Therefore,to follow such an irreversible tendency,more and more new technologies such as deep learning and machine vision have participated in the fish farming,whose patterns also have changed to be more intensive and intelligent.In fish farming,a reasonable amount of feed is the key to saving farming costs and improving economic benefits.So far,artificial feeding and mechanized feeding are still the most 2 widely-applied methods for fish feeding in the chinese market,whose both cause a lot cost and much unnecessary waste due to the amount of feed always depends on the experience of the breeder rather the real hunger situation.In many cases,the amount of feed couldn’t be adjusted appropriately with difficulties getting timely feeding information,which also likely leaded to a water contamination if the feed was superfluous.By analyzing the development of intelligent feeding in fish farming,it could be found that some researches have been used to guide feeding by identifying fish’s instant condition,and even have achieved great results under experimental conditions.But some factors such as environment changes or human activities possibly have bad influences on actual feeding operation.There are still some difficulties in achieving intelligent feeding.In view of the above situation,an intelligent decision feeding system based on machine vision by studying the feeding behavior of surface Micropterus salmoides was designed and tested in this work,which proved that it could realize intelligent feeding of outdoor captive perch.The results and conclusions of this research were as follows:(1)The feeding behavior of Micropterus salmoides was studied.Based on the “zeroemission” captive breeding mode,an above-water video collection platform was built to collect fish feeding videos.By combining feeding regularity and the feeding behavior of fish,it could be found that the initial stage of feeding was more intense and difficult to observe surface feeding and feed residuals,but it is different in the final stage of feeding.In this work the 20 th ～ 40 th s pictures were selected as data samples in each feeding round,and the appetite level of fish was divided into "Strong","Medium","Weak" and "None" 4according to different hunger degree.(2)Fish feeding behavior was recognised based on traditional machine learning.According to the fish appetite level,a fish appetite level image dataset was created,which contained 16,000 images.Through feature engineering totally 17 fish feeding image features including 3 types—image texture,color and shape features were extracted,and then was filtered further to 5 features: Contrast,Homogeneity,Entropy,Correlation and Angular Second Moment(ASM).K-Nearest Neighbors,Support Vector Machine,Random Forest and Stacking traditional machine learning models were constructed and trained based on fish feeding image features.The results showed that Random Forest model had the best recognition effect among the 4 models,whose average accuracy,precision,recall and F1 score were 83.65%、83.67%、83.65% and 83.66%,and the average recognition rate was 18.58 fps.(3)Fish feeding behavior was recognised based on deep learning.The appetite level of fish was recognised by Res Net18,Shuffle Net V2 and Mobile Net V3-Small deep learning models.It was seen that Mobile Net V3-Small was the best model,whose average accuracy,precision,recall and F1 score were 97.10%,97.11%,91.77% and 94.37%,Floating Point Operations was 582.40 MB,parameter amount was 1.53 MB,and the average recognition rate was 32.06 fps.Then,the Mobile Net V3-Small model was improved by adding a dilated convolution layer.Compared with the Mobile Net V3-Small model,the improved Mobile Net V3-Small model improved the average accuracy,precision,recall and F1 score by 0.15%,0.15%,0.41% and 0.28%.In addition,the effects of different learning rates on the recognition performance of the model were explored and compared.The results showed that the average accuracy,precision,recall and F1 score of the improved model trained by Cosine-Warmup learning rate were 98.31%,98.31%,94.82% and 96.53%,which were higher than the optimal fixed learning rate model respectively 0.36%,0.35%,0.72% and0.54%.Finally,the improved Mobile Net V3-Small model trained by the Cosine-Warmup learning rate was thought of as the Micropterus salmoides decision feeding model to recognise the appetite level of fish after comparing different machine learning models(4)Intelligent decision feeding system was designed and tested.The system was divided into application layer,platform layer,transport layer and perception layer according to the composition of the Internet of Things.A small pond feeder was improved—the control system was changed to Node MCU-8266 control system.The test verified that the improved feeder could meet the feeding requirements.In order to realize the information exchange between the feeder and the computer decision system,a Wi-Fi Internet of Things system was built based on MQTT and Alibaba Cloud Internet of Things platform.A system graphical user interface was designed,which could display fish information and video images,recognise the appetite level of fish,and control the feeding of the feeder.Finally,a feeding test for captive perch in the breeding pond was carried out to verify the practicing effect of the intelligent decision feeding system.It could be shown that compared with artificial feeding,the Feed Conversion Ratio of the intelligent decision feeding system decreased by 0.03,and the Weight Gain Ratio increased by 0.11%,while the Specific Growth Rate was the same.Which proved that the overall growth with the intelligent feeding system had an obvious advantage than that of manual feeding.In a conclusion,the system could effectively recognise the appetite level of fish and correspondingly make precise feeding decisions,thereby to replace the breeding staff,and provide a reference for intelligent fish feeding in the outdoor intensive breeding mode. |
PDF Full Text Request