Image classification task is an important branch of computer vision. At present, most of the mainstream CNNs are large in size and take up too much computing resources. The quality-price ratio is not satisfying when the heavy CNNs are used in image classification. So, this work proposes a spatial and channel hybrid attention module (Light-Attention module), an ultra-lightweight but efficient attention module. Given an intermediate feature map, the Light-Attention module will firstly derive the most important attention maps of the channels automatically with global stochastic pooling and Multilayer Perceptron (MLP). Then, the residual structure of Light-Attention module helps to repeatedly introduce global information for secondary screening. For better performance, our pioneering Max module and Mean module to extract key spatial features on the basis of previous operations. In the process of extracting the key attention map, LADNet uses the most labor-saving operation and save a lot of network parameters. The Light-Attention module can be seamlessly integrated into any network, while its parameters and required computing resources are both very small. We verified the excellent performance of the Light-Attention module in extracting key image information through ablation experiments on the Cifar-10, Cifar-100 and ImageNet datasets. For the experiment in Cifar-10, the image classification accuracy achieves 98.7%, achieves 96.5% for the Cifar-100, and for the ImageNet, achieves 83.9%. With the Light-Attention module, LADNet reduces the parameters of convolution neural network to 71% of the original. Compared to the other CNNs used in image classification task, the Light-Attention module uses the least parameters but scores most advanced achievements. The experiments proved that the combination of DRN (Dilated Residual Network) and Light-Attention module achieves superior performance.

Hand pose estimation from RGB images has always been a difficult task, owing to the incompleteness of the depth information. Moon et al. improved the accuracy of hand pose estimation by using a new network, InterNet, through their unique design. Still, the network still has potential for improvement. Based on the architecture of MobileNet v3 and MoGA, we redesigned a feature extractor that introduced the latest achievements in the field of computer vision, such as the ACON activation function and the new attention mechanism module, etc. Using these modules effectively with our network, architecture can better extract global features from an RGB image of the hand, leading to a greater performance improvement compared to InterNet and other similar networks.

Recent research shows that strip pooling can effectively improve the performance of the model. In this paper, we use efficientnet as the backbone network in the structure of UNet to perform strip pooling of feature maps of different scales, and use the multi-head attention mechanism to fuse the obtained features by strip pooling, thus improving the model’s ability to detect remotely sensed roads. Strip pooling can capture the remote dependency of the context and improve the detection ability of slender objects. The multi-head attention mechanism can use the correlation between features of different scales to enhance the expression of their respective features. We apply the proposed model to remote sensing data sets, which can significantly improve the IoU score and the accuracy of pixel classification of the UNet baseline model.

Retinal Image Segmentation is an important part of Medical image. In terms of taking advantage of the characteristics of retinal vessel width changes as well as the local linear feature, in this article, we propose ML-Net, a new model based on U-Net, with the ability to find the linear feature of retinal vessels. ML-Net mainly contains two major blocks: a multi-kernel fusion pooling (MFP) block to preserve spatial information and a linear convolution kernel (LCK) block to capture more strip information. Both blocks are lightweight and can serve as efficient plug-and-play modules in existing scene parsing networks. We apply the proposed ML-Net to three mainstream data sets, namely DRIVE, CHASE-DBI, and STARE. The comprehensive results show that the proposed method outperforms the original U-Net and other state-of-the-art methods for retinal image segmentation.