Research On Biomass Measurement Method Of Aquaculture Juvenile Abalones Based On Deep Learning

Abalone has rich nutritional value and delicious taste.The demand for abalone continues to increase in domestic and foreign markets.However,due to the deterioration of water bodies,illegal fishing and other factors,the natural output of abalone is low,which is far from meeting the market demand for abalone.Therefore,abalone factory farming is developing rapidly.Seed breeding of abalone is an important part of factory farming.During the process of abalone seedling cultivation,it is necessary to strictly control the number of abalone to prevent abalone from growing slowly due to insufficient feed or high breeding density,and low density will affect production efficiency.Abalone is sold as seed in a variety of sizes and needs to be screened and graded.At present,my country relies on manual counting,screening and grading of abalone in the nursery stage,which is a large workload and error-prone.Therefore,it is of great significance to apply convolutional neural network to intelligent detection,counting,screening and grading of juvenile abalone.In this paper,deep learning technology is used to detect the individual position and number of juvenile abalone,recover the complete outline of juvenile abalone and calculate its shell length and circumference,so as to realize the function of real-time monitoring of the number,breeding density and size of juvenile abalone.The main work of this paper has the following three aspects:(1)A detection and counting model FAL＿SSD of juvenile abalone based on improved SSD network is proposed.Under the experimental conditions,a juvenile abalones counting dataset(JAC dataset for short)is constructed.In order to improve the detection accuracy of juvenile abalone with similar color and texture and occlusion,the SSD object detection model is improved.Firstly,a multi-layer feature dynamic fusion method is introduced in the convolutional layer of the SSD network,so that the network can obtain more color and texture information of juvenile abalone and improve the detection accuracy of juvenile abalone under the condition of similar color and texture.Secondly,the multi-scale attention feature extraction method is introduced to eliminate the noise information from the image background,highlight the shape and edge information of the juvenile abalone,and improve the detection accuracy of the juvenile abalone in the case of dense distribution and occlusion.Finally,the loss feedback training method is introduced,and the loss is used as the feedback signal to judge whether to randomly crop and splice the picture,enhance the data diversity,increase the pixels of the juvenile abalone in the picture,and improve the detection accuracy of the smaller juvenile abalone.The AP50 value and AP75 value of the FAL＿SSD model detected on the JAC dataset are 91.14% and 80.14%,respectively,which are superior to the detection results of the SSD model,the FSSD model,the Efficient Det model,the Mutual Guide model and the Varifocal Net model.The accuracy and effectiveness of the FAL＿SSD model proposed in this paper are verified.(2)A recovery and calculation model FR＿Deocclusion of juvenile abalone body shape based on improved U-NET network is proposed.Under the experimental conditions,a juvenile abalones shape calculation dataset(JASC dataset for short)is constructed.In order to recover the complete contour of the individual juvenile abalone and improve the recovery accuracy,the unsupervised de-occlusion recovery model is improved.Firstly,improve the connection mechanism of the U-NET network in the model,deepen the depth of network learning,and improve the accuracy of segmenting adjacent juvenile abalone.Secondly,in view of the problem that the individual recovery accuracy of juvenile abalone is low in the case of similar color,the features of different layers in the network are integrated by superposition,so that the network can obtain more color and edge information of juvenile abalone,and improve the accuracy of complete contour recovery of juvenile abalone.Finally,the m IOU value and p Acc value of the body shape restoration are performed in the JASC dataset using the FR＿Deocclusion model are 86.69% and 89.65%,respectively,which are better than the restoration results of the Self-Supervised Scene De-occlusion model and the ASBU model.The accuracy and effectiveness of the FR＿Deocclusion model proposed in this paper are verified.(3)A biomass measurement system for cultured juvenile abalone is developed.The biomass measurement system of juvenile abalone uses the Windows10 operating system,and the development framework based on pytorch mainly designs 2 interfaces and 8 functions to complete the development of the system.The system includes four modules: login and registration module,detection and counting module of juvenile abalone,calculation module of body shape recovery of juvenile abalone and system database module.Wherein,the gui Main package is used to select pictures in real time,set the interface of the detection counting module and the shape recovery calculation module,and display the visual results of the detection,counting,shape recovery and calculation of the juvenile abalone that in the selected picture.The Single Pic Inter1 package is used to call the weight file of the juvenile abalone detection and counting model to detect and count juvenile abalone,and save the results.The Single Pic Inter2 package is used to call the weight file of the juvenile abalone body shape recovery calculation model to restore and calculate the body shape of juvenile abalone,and save the results.The calculation of the number,individual shell length and circumference of juvenile abalone is realized.