The Research And Application Of Fine-grained Algorithm Based On Deep Learning

Image recognition is one of the most fundamental and important research topics in computer vision.Fine-grained image recognition has become a hot research direction in recent years.Image recognition refers to the processing,analysis and understanding of images by computer to identify objects and targets in various modes,and it further identifies a certain category of objects into different sub-categories.Compared to categories,sub-categories of objects have smaller inter-class differences and larger intra-class differences that are more difficult to distinguish,so fine-grained image recognition is a research topic with more challenging and application value in realistic scenario.Based on the fine-grained image recognition,this paper studies the target detection method based on YOLO(You only look once)and the end-to-end fine-grained image recognition method based on BFCNN(Bilinear fusion convolutional neural network),and applies it to the pig body recognition task.The main contents are as follows:It is a key step to automatically detect and locate the pig body in the image for improving the accuracy and efficiency of the recognition aimed at the huge difference of the data set in the natural scene.With a comparing and analyzing for existing detection algorithm,this paper adopts YOLO methods which is based on end to end and no alternative region for pig body detection.The improvements aimed to the original YOLO are as follows:The network is optimized by using Batch Normalization(BN)operation to accelerate the convergence speed of the network,and the precision of detection is improved by increasing input resolution.The experiment proves that the improved YOLO method has strong robustness and better speed for the detection of pigs and improvement in accuracy.Based on the bilinear deep convolutional neural network BCNN and the idea of feature fusion,this paper proposes a convolutional neural network to achieve fine-grained recognition for the pig images.Unlike the previous idea of fine-grained recognition work using feature extraction neural network for feature extraction and classification,BCNN is an end-to-end fine-grained recognition model,directly taking images as input,via two convolutional neural network-based feature extractors.Then,the extracted features are bilinear operated and aggregated by a pooling function to obtain a more representative image description operator.Through the visual analysis to the feature by a convolutional neural network learning,we found that the features extracted by each convolutional layer of CNN have layered features,the lower layers correspond to the basic features,the higher layers correspond to the features,and Features that combine different layers will have a stronger ability to express.Based on the above observations,this paper adds the coalescent of the front layer features based on BCNN and proposes the BFCNN network.Experiments show that BFCNN has favorable performance in extracting the fine-grained characteristics of pigs,and achieved 98.7%recognition accuracy in the 30 pig data sets.