基于再感知双模型联合训练的散焦模糊检测  

Joint training of dual-model reperception for image defocus blur detection

作　　者：朱智勤 孟骏 李嫄源[1] 齐观秋 李华锋[3] 姚政 ZHU Zhiqin;MENG Jun;LI Yuanyuan;QI Guanqiu;LI Huafeng;YAO Zheng(Chongqing Key Laboratory of Image Cognition,Chongqing University of Posts and Telecommunications,Chongqing 400065,P.R.China;Computer Information Systems Department,Buffalo State College,NY 14222,USA;Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650031,P.R.China)

机构地区：[1]重庆邮电大学图像认知重庆市重点实验室,重庆400065 [2]布法罗州立学院计算机信息系统系,美国纽约14222 [3]昆明理工大学信息工程与自动化学院,昆明650031

出　　处：《重庆邮电大学学报（自然科学版）》2024年第1期86-97,共12页Journal of Chongqing University of Posts and Telecommunications(Natural Science Edition)

基　　金：国家自然科学基金项目(61803061,61906026);重庆市教委重庆市高校创新群体“成渝双城经济圈建设”科技创新项目(KJCXZD2020028);重庆市技术创新与应用发展专项项目(CSTC2019JSCX-ZDZTZX0068)~~。

摘　　要：针对散焦模糊检测(defocus blur etection, DBD)模型训练时没有对响应错误区域学习优化,且在识别过程中部分图像的模糊同质区域和处理边界过渡等位置仍然具有挑战性等问题,提出了再感知双模型联合训练方法和基于注意力机制多尺度语义融合散焦模糊检测网络。将未正确响应的预测区域映射到全新合成图像中驱动模型学习,实现再感知错误位置的图像特征;利用DBD任务的互补性质构建预测网络,组成对焦预测和模糊预测双模型,将互补网络中多余响应的区域反馈到另一个模型上从而提升训练效果;利用多尺度特征融合模块逐渐整合不同尺度的语义信息;在特征提取时设计了全局通道注意力模块,使模型关注预测结果的有效特征信息,增强网络在不同输入场景下的灵活性。在DUT、CUHK和CTCUG数据集上进行的对比实验表明,提出的方法与对比方法中性能最优者相比,F-Measure指标分别提高了0.082、0.051、0.264,MAE指标分别降低了0.032、0.018、0.144。For the defocus blur detection(DBD)model training,there is no learning optimization for the response error area(the area where the extracted image feature information does not correspond to the original image),and the blur of some images is homogeneous during the recognition process.Locations such as regions and handling boundary transitions remain challenging.This paper proposes a re-perceptual dual-model joint training method and a multi-scale semantic fusion defocus blur detection network with channel attention.In the model training phase,the model learning is driven by mapping the incorrectly responded predicted regions to a new synthetic image to re-perceive the image features of the wrong location.This approach involves creating a dual model,comprising a focus prediction model and a defocus prediction model,which leverage the complementary nature of the DBD task.The redundant response areas in one model,which contain excess image feature information,are fed back to the other model to enhance the training effect.Considering the sensitivity of defocus blur features to scale,this paper utilizes a multi-scale feature fusion module to gradually integrate semantic information at different scales.In addition,a global channel attention module is designed during feature extraction to make the model focus on the effective feature information of the prediction results and increase the flexibility of the network under different input scenarios.Comparative experiments show that the F-Measure index of the method in this paper shows improvements of 0.082,0.051,and 0.264 and the MAE index is reduced by 0.032,0.018,and 0.144,respectively.

关 键 词：散焦模糊检测 双模型联合训练 多尺度特征 互补模型 注意力机制 

分 类 号：TN919.8[电子电信—通信与信息系统] TP301.6[电子电信—信息与通信工程]