Houcheng Su

Fully test-time adaptation (FTTA) adapts a model that is trained on a source domain to a target domain during the testing phase, where the two domains follow different distributions and source data is unavailable during the training phase. Existing methods usually adopt entropy minimization to reduce the uncertainty of target prediction results, and improve the FTTA performance accordingly. However, they fail to ensure the diversity in target prediction results. Recent domain adaptation study has shown that maximizing the sum of singular values of prediction results can simultaneously enhance their confidence (discriminability) and diversity. However, during the training phase, larger singular values usually take up a dominant position in loss maximization. This results in the model being more inclined to enhance discriminability for easily distinguishable classes, and the improvement in diversity is insufficiently effective. Furthermore, the adaptation and prediction in FTTA only use data from the current batch, which may lead to the risk of overfitting. To address the aforementioned issues, we propose maximizing the sum of singular values while minimizing their variance. This enables the model's focus toward the smaller singular values, enhancing discriminability between more challenging classes and effectively increasing the diversity of prediction results. Moreover, we incorporate data from the previous batch to realize semantic data augmentation for the current batch, reducing the risk of overfitting. Extensive experiments on benchmark datasets show our proposed approach outperforms some compared state-of-the-art FTTA methods.

The task of domain generalization is to learn a classification model from multiple source domains and generalize it to unknown target domains. The key to domain generalization is learning discriminative domain-invariant features. Invariant representations are achieved using adversarial domain generalization as one of the primary techniques. For example, generative adversarial networks have been widely used, but suffer from the problem of low intra-class diversity, which can lead to poor generalization ability. To address this issue, we propose a new method called auxiliary classifier in adversarial domain generalization (CloCls). CloCls improve the diversity of the source domain by introducing auxiliary classifier. Combining typical task-related losses, e.g., cross-entropy loss for classification and adversarial loss for domain discrimination, our overall goal is to guarantee the learning of condition-invariant features for all source domains while increasing the diversity of source domains. Further, inspired by smoothing optima have improved generalization for supervised learning tasks like classification. We leverage that converging to a smooth minima with respect task loss stabilizes the adversarial training leading to better performance on unseen target domain which can effectively enhances the performance of domain adversarial methods. We have conducted extensive image classification experiments on benchmark datasets in domain generalization, and our model exhibits sufficient generalization ability and outperforms state-of-the-art DG methods.

With the rapid development of computer vision, the application of computer vision to precision farming in animal husbandry is currently a hot research topic. Due to the scale of goose breeding continuing to expand, there are higher requirements for the efficiency of goose farming. To achieve precision animal husbandry and to avoid human influence on breeding, real-time automated monitoring methods have been used in this area. To be specific, on the basis of instance segmentation, the activities of individual geese are accurately detected, counted, and analyzed, which is effective for achieving traceability of the condition of the flock and reducing breeding costs. We trained QueryPNet, an advanced model, which could effectively perform segmentation and extraction of geese flock. Meanwhile, we proposed a novel neck module that improved the feature pyramid structure, making feature fusion more effective for both target detection and instance individual segmentation. At the same time, the number of model parameters was reduced by a rational design. This solution was tested on 639 datasets collected and labeled on specially created free-range goose farms. With the occlusion of vegetation and litters, the accuracies of the target detection and instance segmentation reached 0.963 (mAP@0.5) and 0.963 (mAP@0.5), respectively.

At present, fish farming still uses manual identification methods. With the rapid development of deep learning, the application of computer vision in agriculture and farming to achieve agricultural intelligence has become a current research hotspot. We explored the use of facial recognition in fish. We collected and produced a fish identification dataset with 3412 images and a fish object detection dataset with 2320 images. A rotating box is proposed to detect fish, which avoids the problem where the traditional object detection produces a large number of redundant regions and affects the recognition accuracy. A self-SE module and a fish face recognition network (FFRNet) are proposed to implement the fish face identification task. The experiments proved that our model has an accuracy rate of over 90% and an FPS of 200.

Efficient and accurate extraction of water areas from remote sensing images is a popular research topic. Currently, researchers have attempted to use neural networks to extract water from remote sensing images. However, most of these studies used computer vision techniques to improve the model results without considering the multi-band information unique to remote sensing images. Thus our study proposes an improved DeepLabv3+ network to increase the water body extraction accuracy in urban remote sensing images. The DeepLabv3+ network has the characteristics of extracting image features at multiple scales. We improved the network structure to incorporate multi-band features. By comparing several multi-band input methods, the feature map calculated by the normalized difference water index (NDWI) was converted into an input suitable for the neural network by comparing several multi-band input methods. Simultaneously, we developed a parallel convolution structure to combine the NDWI feature map with a standard false color remote sensing image during feature extraction. This allows the network to focus more on image areas that may be water bodies. We used atmospherically corrected Sentinel-2A L2A-level data, divided the training set at multiple scales, and conducted several experiments. The results show that the proposed network can improve water extraction accuracy when training subregions are unified from different sizes to 512 × 512. Finally, we used the model to extract water bodies from remote sensing images from different regions. We combined the images with visual interpretation to verify the reliability of the model results. Moreover, the model scores of four types of multi-scale neural networks in the two categories are compared, which proves the effectiveness of the method.