The spatial distribution of remote-sensing scene images is highly complex in character, so how to extract local key semantic information and discriminative features is the key to making it possible to classify accurately. However, most of the existing convolutional neural network (CNN) models tend to have global feature representations and lose the shallow features. In addition, when the network is too deep, gradient disappearance and overfitting tend to occur. To solve these problems, a lightweight, multi-instance CNN model for remote sensing scene classification is proposed in this paper: MILRDA. In the instance extraction and classifier part, more discriminative features are extracted by the constructed residual dense attention block (RDAB) while retaining shallow features. Then, the extracted features are transformed into instance-level vectors and the local information associated with bag-level labels is highlighted by the proposed channel-attention-based multi-instance pooling, while suppressing the weights of useless objects or backgrounds. Finally, the network is constrained by the cross-entropy loss function to output the final prediction results. The experimental results on four public datasets show that our proposed method can achieve comparable results to other state-of-the-art methods. Moreover, the visualization of feature maps shows that MILRDA can find more effective features.

Capsule network is a promising model in computer vision. It has achieved excellent results on simple datasets such as MNIST, but the performance deteriorates as data becomes complicated. In order to address this issue, we propose a deep capsule network in this paper. To deepen the capsule network, we present a new tensor capsule based routing algorithm and the corresponding convolution operation. Compared to vector capsules, tensor capsules can capture more instance-level information. Together, the relevant convolution operation is beneficial for reducing the amount of parameters in the routing process. Furthermore, we propose a dropout mechanism for vectors and tensors in order to alleviate the potential overfitting problem. Finally, we also inject the multi-scale capsules of the middle layers into a multi-scale decoder to pursue more details of an image and more clear reconstructed image. Experimental results on CIFAR10, Fashion-MNIST, and SVHN demonstrate that the proposed deep tensor network can achieve very competitive performance compared to other state-of-the-art capsule networks.

Retinal images are the most direct and effective basis for Diabetic Retinopathy (DR) diagnosis. With the rapid development of deep learning, the technology of retinal image-assisted diagnosis based on deep learning is widely used in the field of DR intelligent diagnosis. However, the training of deep neural network usually requires a large number of annotated samples, but retinal images annotated by professional doctors are cost-expensive and difficult to obtain, which limits the application of deep learning technology in DR intelligent diagnosis. In order to alleviate the scarcity of labelled retinal images, we propose an end-to-end conditional generative adversarial network with class feature loss and improved retinal detail loss. The network combines the above two losses with the adversarial loss, and jointly constrains the generator to generate high-quality retinal images. The proposed retinal detail loss is summed over physiological detail loss which is meant to preserve high-level semantic features of the physiological details contained in the fundus images and pixel loss which ensures the low-level features in synthesized image will not deviate from the real image. In addition, the class feature loss constrains the synthesized images to be consistent with the real images in class features representation, which further makes the synthesized images have pathological features of the corresponding grade. The generated images by the proposed network are evaluated from three objective metrics including the subjective effect and the FID, SWD, which are used to evaluate the quality and diversity of generated images, and the effect of retinal vessel segmentation, respectively. Experimental results demonstrate that our synthesized images have superior performance on both the quality and quantity.

The rapid increase in spatial resolution of remote sensing scene images (RSIs) has led to a concomitant increase in the complexity of the spatial contextual information contained therein. The coexistence of numerous smaller features makes it challenging to accurately locate and mine these features, which in turn makes accurate interpretation difficult. In order to address the aforementioned issues, this paper proposes a dynamic convolution covariance network (ODFMN) based on omni-dimensional dynamic convolution (ODConv), which can extract multi-dimensional and multi-scale features from RSIs and perform statistical higher-order representation of feature information. First, in order to fully exploit the complex spatial context information of RSIs and at the same time improve the limitation of a single static convolution kernel for feature extraction, we constructed a omni-dimensional feature extraction module (ODFE) based on dynamic convolution, which fully extracts the four-dimensional information within the convolution kernel. Then, to make full use of the full-dimensional feature information extracted from each level in the network, the feature representation is enriched by constructing multi-scale feature fusion module (MFM) to establish relationships from local to global. Finally, higher-order statistical information is employed to address the challenge of representing first-order information for smaller object features, which is inherently difficult to do. Experiments conducted on publicly available datasets have demonstrated that the method achieves high classification accuracies of 99.04%, 95.34%, and 92.50%, respectively. Furthermore, the method has been verified to have high capture accuracy for feature target contours, shapes, and spatial context information through feature visualisation.

Synthetic aperture radar (SAR) image classification is an important task in remote sensing applications. However, it is challenging due to the speckle embedding in SAR imaging, which significantly degrades the classification performance. To address this issue, a new SAR image classification framework based on multi-feature fusion and adaptive kernel combination is proposed in this paper. Expressing pixel similarity by non-negative logarithmic likelihood difference, the generalized neighborhoods are newly defined. The adaptive kernel combination is designed on them to dynamically explore multi-feature information that is robust to speckle noise. Then, local consistency optimization is further applied to enhance label spatial smoothness during classification. By simultaneously utilizing adaptive kernel combination and local consistency optimization for the first time, the texture feature information, context information within features, generalized spatial information between features, and complementary information among features is fully integrated to ensure accurate and smooth classification. Compared with several state-of-the-art methods on synthetic and real SAR images, the proposed method demonstrates better performance in visual effects and classification quality, as the image edges and details are better preserved according to the experimental results.