Fine-grained Fish Classification Based On Bilinear Network

Effective classification of various fish categories under water has great practical and theoretical significance.Since the living environment is highly complex,fish images are greatly affected by light,angle,posture,etc.Also,they have properties of high inter-class similarity and large intra-class variety.These factors make fish classification a challenging task.To cope with these challenges,three different fish images classification models based on bilinear network are proposed,which all have good performances.(1)A fine-grained fish classification model based on spatial transformation bilinear network is proposed.It consists of a spatial transformer network and a bilinear network.Specifically,spatial transformer network works as an attention mechanism to remove the complex background and selecting the region of interest in the image,and the bilinear network extracts the bilinear features of the images,which respond to the discriminative part of the target.The model can be trained in an end-to-end way.Experiments on two datasets verify the effectiveness and superiority of the proposed algorithm.(2)A fish classification network based on selective convolutional descriptor is proposed.The network is similar to the method(1)in that it selects the target area in the image first and then uses the bilinear network to classify,but the method used to extract the target area is different.The model firstly locates the target object in the image based on the selective convolutional descriptor and the other parts are treated as background and noise.This localization is unsupervised,without utilizing bounding boxes,image labels,object proposals,or additional learning.(3)Aiming at the low resolution of fish images under real environment,a network combining super resolution and classification is proposed.First,an advanced image super-resolution technique is used to enrich the details of the image that are important for classification,and then classify.Extensive experiments have shown that the proposed three algorithms have state-of-the-art performances.