Research On Image Classification Based On Bilinear Convolutional Neural Network

Bilinear convolutional neural network is a fine-grained image classification problem in image classification tasks.Fine-grained image classification is an important research direction in the field of computer vision.The purpose of this task is to divide sub-categories in coarse-grained images.It plays an important role in the application and protection of the ecological species,and therefore there are broad prospects for research and development.Since the difference between sub-category objects is usually subtle in fine-grained images,it is generally only possible to rely on small local differences to complete the classification task.How to better construct a fine-grained image classification model has become the focus of this task.In recent years,the deep convolutional neural network has developed rapidly and demonstrated strong feature learning ability,which has made the deep learning method widely concerned and in-depth research in the field of image classification.Based on the classical fine-grained image classification method,this paper improves the method from three different angles to solve the different problems in the fine-grained image classification task.The main research contents and innovation work of this paper are as follows:In the fine-grained image classification,the use of the bounding-box is an important factor,which plays a important role in removing irrelevant background interference.In the existing fine-grained image classification method,the bounding-boxes used are all provided by the official dataset.Therefore,it is impossible to accurately locate objects in an image for a fine-grained image dataset that does not provide a bounding-box.Based on the classical fine-grained image classification method,this paper adds the detected annotation information,and proposes a fine-grained image classification method based on object detection for the first time.The bounding-box provided by the officical dataset is no longer used in this method.Instead,the object detection algorithm is used to mark the foreground objects in the images,then the detected bounding-box is used in the network training process.This not only reduces the amount of work caused by manual annotation of images,but also expands the application to other datasets of various categories,effectively improving the practicability of annotation information in fine-grained image classification algorithms.In addition,this paper also evaluates the use of the bounding-box,so that the detected information can be used to the maximum extent.In the process of training the network,high-dimensional features often contain more image information,which is helpful for the specific expression of fine-grained images.But the high dimension will take up more storage space,consume more memory to run,introducing excessive amount of parameters during network training.Therefore,to solve this problem,this paper proposes a compact second-order deep convolution feature,which uses the dot product to calculate the second-order feature representation of modeled image information.At the same time,based on this feature modeling,three new network architectures are designed,which can control the size of the feature vector expression dimension and the number of parameter quantities in network training without losing the image-specific information,getting competitive classification accuracy in fine-grained image classification task.Modeling of second-order statistical features is a common feature modeling method in fine-grained image classification tasks.For the second-order information of the image,it can be modeled from multiple angles such as spatial correlation and channel correlation of features.For the role of these second-order information in promoting mutual expression in network training,this paper proposes a feature fusion research method based on second-order statistical information.At the expression level in the image,the image of the second-order features of different modeling method for fusion were obtained.The two feature modeling methods of fusion are: the compact second-order depth convolution feature proposed in the third chapter of this paper,and the second-order spatial covariance feature of the image.The fusion method achieves a higher performance improvement in the fine-grained image classification task,and proves that there is complementarity between the second-order statistical information obtained by different methods.