摘要
对网站图像敏感信息识别检测问题,提出一种基于改进深度学习的图像敏感信息识别方法。通过特征融合网络,将经全局优化后的区域检测网络与全局识别网络提取特征相融合,并引入注意力机制,对图像中包含敏感部位的区域进行加权聚焦,提高模型检测效率和准确性。实验证明,采用通过全局特征优化后的区域检测网络,平均检测精度提高1%,相较于SSD、Faster R-CNN等目标生成网络,平均检测精度高8.54%与10.63%,提取结果更精准;融合局部特征的全局识别网络,识别精度随着局部特征提取准确度上升而上升,当提取种类到达10种时,识别精度比未加入高1.8%;通过引入注意力机制,本模型检测准确率提升明显,当聚焦点数为7时,比未引入高0.7%;最终,相较于未包含局部特征的ResNet50网络、虽然未包含局部特征但结构更复杂的ResNet101网络,与虽然考虑局部特征,但未与全局特征进行融合的DMCNet网络,本模型检测准确率平均高出3.25%、2.15%和6%,且耗费时间较短,具有更高的鉴别力和检测效率。
To solve the problem of website image sensitive information recognition,this paper proposes an improved deep learning based image sensitive information recognition method.The feature fusion network is used to integrate the extracted features of the globally optimized regional detection network and the global recognition network.Moreover,the attention mechanism is introduced to focus the sensitive parts in the image,so as to improve the efficiency and accuracy of model detection.Experiments show that the average detection accuracy is improved by 1%by using the regional detection network optimized by global features.Compared with the network generated by SSD and Faster R-CNN,the average detection accuracy is higher by 8.54%and 10.63%,and the extraction results are more accurate.When the global recognition network of local features was integrated,the recognition accuracy increased with the increase of local feature extraction accuracy.When the extraction types reached 10,the recognition accuracy was 1.8%higher than that without the addition of local features.By introducing the attention mechanism,the accuracy of model detection in this paper is significantly improved.When the number of focus points is 7,it is 0.7%higher than that without the introduction.Finally,compared with ResNet50 network which does not contain local features and DMCNet network which considers local features but does not integrate global features,the detection accuracy of the model in this paper is 3.25%and 6%higher on average,and the time consuming is shorter,with higher discrimination and detection efficiency.
作者
李选臣
LI Xuanchen(Shaanxi Institute Of Mechatronic Technology,Baoji Shaanxi 721001,China)
出处
《自动化与仪器仪表》
2023年第10期36-39,44,共5页
Automation & Instrumentation
关键词
图像识别
区域检测
全局分类
注意力机制
特征融合
image recognition
area detection
global classification
attention mechanism
feature fusion
