Analysis And Research Of Key Technologies For Fine-grained Image Recognition Based On Convolutional Neural Networks

Image recognition can be divided into generic and fine-grained image recognition.Compared with the general image recognition that only realizes the obvious category,fine-grained image recognition classifies the object into a fine dub-class category.Convolutional neural networks(CNNs)use multiple processing layers that contain complex structures or are composed of multiple nonlinear transformations to realize the learning of big data,and achieved far better results than traditional methods.Through learning a large number of data,they have achieved far better results than traditional algorithms,and have become the mainstream technology in the field of fine-grained image recognition.Methods based on CNNs have become the main technology in fine-grained image recognition.However,since the actual scene of the recognition image contains various complex factors,such as illumination changes,occlusion,and target movement,the sub-class categories invariably own the same global appearance,and intra-class image features have large differences.Accurately distinguishing fine-grained categories is still an exceedingly further challenge.In this paper,we take fine-grained image combined with deep neural networks,to conduct theoretical analysis,method research,implementation,and verification of key technologies such as data augment,method for fine-grained image recognition under simple background,method for fine-grained image recognition,and fine-grained image detection.The main research contents are as follows.1)A data augmentation method for fine-grained image recognition is presented.Most of the current data augmentation methods are designed for general image recognition.When that method is directly applied to fine-grained image recognition,the improvement of fine-grained recognition performance is very limited.we construct an oriented pair interaction mixing for augmenting fine-grained image recognition,which can improve the performance of the fine-grained recognition model without adding model parameters.The core idea is to measure the similarity of two image samples in the feature space through the Euclidean distance,find the images with the smallest similarity in the intra-class and the largest similarity in the inter-class,respectively mix them to generate new images.The multi-sample image blending augmentation improves the generalization ability of the recognition model and the ability to discriminate key feature regions.Experiments on three large public datasets verify the effectiveness and rationality of the proposed algorithm.2)An input-aware and probabilistic prediction convolutional neural network is presented.For the fine-grained image recognition under simple background,background information has little effect on the recognition result,but the sub-class categories invariably own the same global appearance,and intra-class image features have large differences,which makes it difficult to recognize such images with high accuracy.We proposed an input-aware and probabilistic prediction convolutional neural network,which including an one input-aware module and one probabilistic prediction module.Input-aware module develops raw images automatically into the global-scale image,the object-scale image and the part-scale image,by introducing an attention mechanism.Probabilistic prediction module uses three dynamic probabilistic parameters to estimate the prediction of each CNN branch separately,and then combined the three CNN votes for the final decision.Evaluation results from a large dataset of patients showed that the proposed IAPP-CNN achieved the competitive performance for the chromosome recognition task,surpassing the performance of a competitive baseline created by state-of-the-art methods.3)A two-level progressive attention convolutional network is presented.In a complex background,background information has a significant impact on the recognition results.When recognizing fine-grained images under complex backgrounds,the locating and learning of discriminative features is the key for fine-grained image recognition.To solve the problem that the difficulty of fine-grained feature positioning under complex background leads to insufficient learning and affects the final recognition performance.We proposed a two-level progressive attention convolutional network,which including backbone network,MCAF module,and CEA module.Input an image into backbone feature extraction network,we can obtained a feaute map.The MCAF module is used to find distinctive feature map channels which significantly responds to specific regions.The CEA module is further assign weight values to feature map elements.Compared to previous models basing on attention mechanism,the model can extract non-correlated part features which spread over object foreground areas,further improving the recognition accuracy.Experimental results on datasets demonstrate that the method achieves competitive performance.4)A cascade fine-grained image detection method based on joint optimization is proposed.In practical production applications,it is very common for an image to contain multiple different kinds of targets.As an extended task of fine-grained image recognition,fine-grained image detection can locate multiple targets and recognize the fine-grained category of each object.When the general image detection algorithm is directly applied to fine-grained image detection,due to the characteristics of small inter-class difference and large intra-class difference in fine-grained image,it will lead to the problem of accurate target positioning but inaccurate recognition,which seriously restricts the improvement of detection accuracy.In order to solve the problem of inaccurate fine-grained image recognition in fine-grained image detection task,a jointly optimized cascade fine-grained image detection algorithm is proposed.The algorithm integrates the data augmentation algorithm and fine-grained image recognition algorithm proposed in the previous chapters,designs a novel strategy of joint data optimization and joint prediction optimization to improve the detection rate of fine-grained image,and focuses on solving the problems of image mismatch and inaccurate recognition in network cascade.Experimental results on public datasets show that this method can effectively improve the accuracy of fine-grained image detection.