The Study On Classification And Recognition Of Shellfish Images Based On Densely Connected Neural Network

The shellfish aquaculture industry is booming in our country.Nowadays,the demand for shellfish farming products is also increasing not only in coastal cities,but also in inland cities.In many cases,people want to know more shellfish but it is difficult to accurately identify them.However,there are many types of shellfish and there are many similar shellfish.Although the application field and scope of machine vision technology has been very extensive,there are few studies on its application in the field of shellfish classification and recognition.And related shellfish image data is also difficult to obtain,so we constructed a shellfish image database and used the database in the classification and recognition task of shellfish images.In order to improve the recognition accuracy of shellfish images of different categories with similar features and reduce the model complexity as much as possible,we introduced multi-scale features into the densely connected neural network and proposed the improved method for multi-scale perception of densely connected neural network.At the same time,new weight ratio pooling rules are proposed for further constraints.In order to improve the classification effect of the model,we added a center loss function constraint on the above basis,and further optimizes the geometric center pruning of the network.Thus,the classification and recognition of shellfish images of similar categories are realized.Experimental results show that the method in this paper achieved high-accuracy shellfish image recognition,effectively reduce the number of training parameters required,fast convergence,strong robustness,strong anti-interference ability.Especially for different types of shellfish with similar morphological characteristics,the classification and recognition effect is remarkable.The specific work and innovations are as follows:(1)In view of the fact that there is no relevant shellfish image database in China,and there are few shellfish categories in related research.We constructed a shellfish image database containing 33988 images in 9 categories,which includes Glossaulax Didyma,Rapana Peichiliensis and Buccinium Perryi with large differences in morphological,and other six species of shellfish with small differences in morphology and characteristics,which are Dosinia Laminata and Japanese Dosinia,Solen Strictusand Solen Gouldi,Scapharca Subcrenata and Scapharca Broughtoniia.In order to further expand the shellfish image database,we used the DCGAN to augment the original shellfish image database,establish a shell image database,and apply the database to the subsequent shell classification.(2)In view of the peculiarity of the growth time and environment of shellfish images,that is,there are certain differences between shellfish in the same category and small differences from different shellfish,we introduced the idea of multi-scale.Based on the densely connected neural network,a multi-scale perception convolution module is constructed.First,the convolution features are grouped,and then operated in different convolution channels,and the above-mentioned convolution features are multi-scaled by cascading.Perception,on this basis,a new weight ratio pooling rule is proposed,which gives the extracted features different weight ratios according to size,avoiding the discarding of too many features by the maximum pooling rule.The cancellation of the positive and negative features in the average pooling rule.The situation makes the extracted features more comprehensive,enhances the anti-interference of the pooling process,and improves the accuracy of shellfish recognition with similar morphological features.(3)In order to further improve the recognition accuracy of shellfish images with similar morphological features and reduce the model complexity as much as possible,we added a center loss function to the densely connected neural network with multi-scale perception to constrain.We not only considered the intra-class distance,but also considered the distance between classes,which improved the classification accuracy and effect.Due to the large amount of classification data for training and in order to shorten the training time,reduce the amount of calculation and increase the rate,we also introduced the geometric center pruning algorithm for optimization.In order to reduce the complexity of the model and take into account the accuracy of image classification,we changed the training method of the model.The method used a precision recovery method after pruning.It can not only effectively eliminate redundant features without destroying the original network model architecture,but also realized a linear combination of more features and cooperate with dense.The dense connection characteristics of the neural network transfer effective features to deeperlayers,so that the resulting network structure has some flexibility,and retains as many important features as possible,removing secondary features,thereby improving image classification accuracy.Experimental results shows that the method in this paper can achieve high-accuracy shellfish image recognition,effectively reduce the number of training parameters required,fast convergence speed,and strong robustness.Especially for different types of shellfish images with similar morphological characteristics,the classification and recognition effect is remarkable.