Study On Deep Learning Application Of Fishery Image Recognition

People's demands for the quality of life are getting increasingly higher with the development of society nowadays.However,aquatic products,as a major part of people's daily food,have revealed a lot of safety and supply problems in recent years,and have attracted great attention.The aquaculture industry scale has rapidly expanded since China established the principle of “Taking culture as the main way”.Meanwhile,the industry is very dependent on the professionalism of employees due to its long history,it usually takes a number of years to cultivate sufficient professional skills of new employees to perform their works,hence this dependence has caused much difficulty to build new aquaculture bases.On the other hand,the traditional culturing and management methods have a large amount of work that is performed by manual processing with poor efficiency,which leads to unclear evaluation standards,and has seriously restricted the industry's development in the direction of high efficiency and precision.These problems exposed a large number of shortcomings under the increasing demand for production and safety nowadays.Deep Neural Network(DNN)originates from Artificial Neural Network(ANN),which is based on the Multi-Layer Perceptron(MLP)proposed by Geoffrey Hinton in 1986.Since Hinton research group's image recognition model “AlexNet” won the champion in the ImageNet competition with a huge advantage over the second place(using SVM method)in 2012,deep learning has become a research hotspot.Now the most effective way of deep learning application in the field of image processing is Deep Convolutional Neural Network(DCNN),which performs feature extraction on input image with a set of small scale kernels by multiple convolutional sampling.Usually the features of lower layers are abstract details of the input image such as textures,edges,while the kernels of higher layers have larger receptive field,so the extracted features of these kernels are more specific and identifiable.DCNN abandoned the step of feature design,instead,the convolutional network extracts features completely automatically,and current image recognition model based on DCNN has already surpassed human's image recognition capabilities.This topic combined with the development of deep learning,integrate intelligent technology into the aquaculture industry by using deep learning technology to design an aquatic animal image recognition model for the aquaculture step,and an object detection and recognition model for management.The main content of this paper can be summarized as the following aspects:(1)Aiming at the difficulty of develop a DCNN model for aquatic animal image recognition from scratch,a method based on the parameter transfer strategy using fine-tune to retrain the pre-trained model was proposed.Firstly,the image was normalized and pre-processed by data augmentation.Furthermore,the weights of high-level convolution modules were set to trainable to be adaptive on the basis of modifying source model's fully connected step classification layer.Finally,the performance experiments were conducted on various network structures and different proportion of trainable parameters using training time and recognition accuracy on validation set as the evaluation benchmark.The experimental results demonstrated that the highest retrained model's classification accuracy could reach 97.4%,which were 20 percentage points higher than the source model,and the ideal performance could be obtained when the proportion of trainable parameters was around 75%.It is proved that the fine-tune method can effectively obtain a deep neural network image classification model with good performance under low-cost development condition.(2)Aiming at the shortcomings of traditional object detection and recognition methods,such as the low precision and limitations of manual feature design,this paper chose the Single Shot Multi-Box Detector(SSD)as basic algorithm,implemented a method to modify the pre-trained SSD300 model by adopting the Inception structure to replace the extra convolution layers in SSD300 after VGG16 structure.The originally unified 3×3 kernels ware replaced by 1×1,3×3,5×5(which was split into double 3×3 kernels),these multi-scale kernels could extract more feature information without increasing the model complexity.The experimental results showed that this modified model could reach 79.4% mean average precision(mAP)on the VOC2007 test set.Compare with the original SSD300 model,the accuracy of the detection of small objects is significantly improved,about 4~9 percentage points,and about 1 percentage point improvement in the mAP.The research results of this paper provided the theoretical and practical basis for culturing and daily management of the aquaculture industry,which also proposed a method for balance the performance of model and development cost,and proved that deep learning,this emerging intelligent technology could be more commonly used in real situations and could promote the combination of modern technology and traditional industries with a wide range of application prospects.