Research And Application Of Feature Enhancement Technology In Convolutional Neural Network

In recent years,thanks to the powerful feature extraction capability of deep con-volutional neural networks,it has been widely used in the field of image processing.The endless design of new network structures further expands the boundary of convo-lutional neural network feature expression capabilities.From the perspective of feature enhancement,this paper studies how to enhance the feature extraction capabilities of the network without changing the basic network structure.Firstly,from the perspective of the attention mechanism to enhance the ability of convolutional neural networks to extract features,a feature difference model（FD-model）is proposed to enhance the feature expression ability of convolutional neural networks.Existing attention models calculate the attention value based on the current feature distribution that needs to be applied to the current feature.Because the low layers of convolutional neural networks extract common features,it is difficult for those attention model to achieve good results in the low layers of the network.To this end,this paper proposes a new lightweight attention unit,feature difference model.It uses the difference between the two feature maps to generate the attention mask..Experiments on four public benchmark datasets show that the FD model can help improve the performance of deep convolutional neural networks.In particular,Res Net44（6.10%error rate）with the FD model obtains better results than Res Net56（6.24%error rate）,while the former has 29%less network parameters than the latter.Secondly,from the perspective of feature fusion,this paper designs an algorithm to improve the feature expression ability of convolutional neural networks,and proposes a bidirectional feature aggregation network（Bi FAN）.Feature pyramid networks have achieved impressive results in the field of object detection and instance segmentation by aggregating features of different scales,especially in the detection of small objects.However,for large objects with a large proportion of background information in the bounding box,the feature pyramid network is difficult to obtain good results.Based on this,this paper proposes a bidirectional feature aggregation network,which is mainly composed of multi-scale feature fusion module,top-down feature aggregation module and bottom-up feature aggregation module.This two-way network structure can greatly enrich high-level semantic features,which improving the ability of object detection.Experiments on the COCO dataset show that this method has achieved the best results on metrics of₅₀and.Finally,this paper applies feature enhancement technology to the recognition and classification of cervical cancer images.Colposcopy screening is a cheap and easily available medical technique and is one of the important methods for the early diagnosis of cervical cancer.This paper proposes a cervical cancer image recognition scheme based on the feature enhancement technology proposed above to assist in the early screening of cervical cancer.It is mainly composed of two parts:the extraction of the diseased tissue in the cervical cancer image and the classification of those diseased tissue.In this paper,Bi FAN is used to extract the diseased tissue in the image,and then a lightweight attention convolutional neural network is designed to extract the feature of the diseased tissue and classify it.In particular,the introduced spatial attention mechanism allows the neural network to pay more attention to the diseased tissue in the image.Experimental results show that the method has good recognition ability for cervical cancer images,and the accuracy rate is 68.03%,which is 6%higher than other methods.