Five papers recommended for the new development of convolutional neural network in deep learning

1、Deformable CNN and Imbalance-Aware Feature Learning for Singing Technique Classification

Yuya Yamamoto, Juhan Nam, Hiroko Terasawa

https://arxiv.org/pdf/2206.12230

Singing technology is the use of timbre 、 The pitch and other components of the sound fluctuate in time to perform an expressive vocal performance . Their classification is a challenging task , This is mainly due to two factors :1) There are many kinds of fluctuations in singing skills , Affected by many factors ;2) Existing datasets are unbalanced . To solve these problems , This paper develops a new audio feature learning method based on deformation convolution , The class weighted loss function is used to decouple the feature extractor and the classifier . Experimental results show that :1) Deformable convolution improves the classification effect , Especially when applied to the last two layers of convolution ;2) Retrain the classifier , And through the smooth inverse frequency weighted cross entropy loss function , Improved classification performance .

2、CNN-based fully automatic wrist cartilage volume quantification in MR Image

Nikita Vladimirov, Ekaterina Brui, Anatoliy Levchuk, Vladimir Fokin, Aleksandr Efimtcev, David Bendahan

https://arxiv.org/pdf/2206.11127

The detection of cartilage loss is very important for the diagnosis of osteoarthritis and rheumatoid arthritis . So far, there have been a large number of papers on the automatic segmentation tool for cartilage evaluation of joint magnetic resonance images . Compared with knee or hip , The structure of carpal cartilage is more complex , Automatic tools developed for large joints cannot be used for the segmentation of wrist cartilage . However, the automatic wrist cartilage segmentation method has high clinical value , So the paper through optimization U-Net Depth of architecture and increased attention layer (U-Net_AL), Assessed U-Net The performance of four excellent variants of the architecture . and · Compare the corresponding results with the previous design based on patch Convolution neural network of (CNN) Compare the results of . The evaluation of segmentation quality is based on the use of several morphologies （2D DSC、3D DSC、 precision ） And the comparative analysis of volume index and manual segmentation . These four networks are better than those based on patch Of CNN.U-Net_AL Calculated three-dimensional DSC The median (0.817) Significantly larger than the corresponding three-dimensional calculation of other networks DSC value . Besides U-Net_AL CNN Provides minimum average volume error (17%) And the highest relative to the real value Pearson The correlation coefficient (0.765). Use U-Net_AL The reproducibility of calculation is greater than that of manual segmentation . With an additional attention layer U-net Convolutional neural network provides the best performance of wrist cartilage segmentation . For use under clinical conditions , The trained network can fine tune the data set representing a specific group of patients .

3、EdgeNeXt: Efficiently Amalgamated CNN-Transformer Architecture for Mobile Vision Applications

Muhammad Maaz, Abdelrahman Shaker, Hisham Cholakkal, Salman Khan, Syed Waqas Zamir, Rao Muhammad Anwer, Fahad Shahbaz Khan

https://arxiv.org/pdf/2206.10589

In order to pursue the continuous improvement of accuracy , It is usually necessary to develop large and complex neural networks . This model requires high computational resources , Therefore, it cannot be deployed on edge devices . Therefore, building a resource efficient general network has received great attention in many application fields . The work of this paper effectively combines CNN and Transformer The advantages of the model , And propose a new efficient hybrid architecture EdgeNeXt. Especially in EdgeNeXt in , Introduced segmentation depth transpose attention (SDTA) Encoder , The encoder divides the input into multiple channel groups , We use deep convolution and cross channel self attention to implicitly increase the receiving field and encode multi-scale features . By sorting in 、 A large number of experiments on detection and segmentation tasks , The advantages of the proposed method are proved ： Relatively low calculation requirements and superior to the most advanced methods . With 2.2% The absolute gain sum of 28% Of FLOP The reduced absolute gain exceeds MobileViT.EdgeNeXt The model has 5.6M Parameters , stay ImageNet-1K It has been realized. 79.4% top-1 precision .

4、Scaling up Kernels in 3D CNNs

Yukang Chen, Jianhui Liu, Xiaojuan Qi, Xiangyu Zhang, Jian Sun, Jiaya Jia

https://arxiv.org/abs/2206.10555

2D CNN and ViT The latest developments in the field show that , A large kernel is essential for sufficient receptive fields and high performance . Inspired by these papers , Change the paper to study 3D The feasibility and challenges of large kernel design . Prove that 3D CNN There are more difficulties in performance and efficiency in the application of large convolution cores . stay 2D CNN Existing technologies that work well in 3D It is invalid in the network , Including popular depth-wise convolutions. In order to overcome these problems , This paper proposes the space group convolution and its large kernel module （SW-LK block ）. Avoid the traditional 3D Optimization and efficiency of large kernel . The big core proposed in this paper 3D CNN The Internet , namely LargeKernel3D, For all kinds of 3D Mission （ Including semantic segmentation and object detection ） Significant improvements have been made . It's in ScanNetv2 Semantic segmentation realizes 73.9% Of mIoU, stay NDS nuScenes Based on the object detection benchmark 72.8%, stay nuScenes LIDAR Number one on the charts . Through simple multimodal fusion ,NDS Further upgrade to 74.2%.LargeKernel3D Got its CNN and Transformer An equivalent or better result . And it is proved for the first time that a large kernel is suitable for 3D Network is feasible and essential .

5、MEStereo-Du2CNN: A Novel Dual Channel CNN for Learning Robust Depth Estimates from Multi-exposure Stereo Images for HDR 3D Applications

Rohit Choudhary, Mansi Sharma, Uma T V, Rithvik Anil

https://arxiv.org/pdf/2206.10375

In display technology HDR Capture 、 Processing and displaying solutions is critical . Developing cost-effective 3D HDR In video content , Depth estimation of multi exposure stereo image sequences is an essential task . This paper proposes and develops a new depth architecture for stereo depth estimation of multiple exposures . The proposed architecture has two new components .1、 The stereo matching technology used in traditional stereo depth estimation is improved , Deployed a three-dimensional transfer learning method , This method avoids the requirement of cost volume construction , The adoption is based on ResNet Double encoders and single decoders with different weights CNN Feature fusion , And use the effentnet To learn parallax .2、 Using robust parallax feature fusion method , The parallax maps of stereo images under different exposure levels are combined . The parallax images obtained by different exposures are combined and the weight images are used to calculate different qualities . The final predicted disparity map has stronger robustness , And retain the best feature of depth discontinuity . Proposed by the paper CNN Architecture in challenging scene streams and different exposures Middlebury On the stereo data set , It surpasses the most advanced monocular and stereo depth estimation methods in both quantitative and qualitative aspects . The architecture also performs very well in complex natural scenes , It proves that it can be applied to all kinds of 3D HDR Applications are useful .

https://avoid.overfit.cn/post/518cdba7f3174604bb8236cee180e353

author ：monodeep