In recent years, deep learning has greatly improved the ability of wheatear detection. However, there are still three main problems in wheatear detection based on unmanned aerial vehicle (UAV) platforms. First, dense wheat plants often overlap, and the wind direction will blur the pictures, which obviously interferes with the detection of wheatears; second, due to the different maturity, color, genotype, and head orientation, the appearance will also be different; third, UAV needs to take images in the field and conduct realtime detection, which requires the embedded module to detect wheatears quickly and accurately. Given the above problems, we studied and improved YoloV4, and proposed a robust method for wheatear detection using UAV in natural scenes. For the first problem, we modified the network structure, deleted the feature map with a size of 19×19, and used k-means algorithm to re-cluster the anchors, and we proposed a method of prediction box fusion. For the second problem, we used the pseudo-labeling method and data augmentation methods to improve the generalization ability of the model. For the third problem, we simplified the network structure, replaced the original network convolution with the improved depthwise separable convolution, and proposed an adaptive ReLU activation function to reduce the amount of calculation and speed up the calculation. The experimental results showed that our method can effectively mark the bounding of wheatears. In test sets, our method achieves 96.71% in f1-score, which is 9.61% higher than the state of the art method, and the detection speed is 23% faster than the original method. It can be concluded that our method can effectively solve the problems of wheatear detection based on the UAV platform in natural scenes.
With the development of deep learning, person re-identification (ReID) has been widely concerned and studied. At present, in practical application, there are three main problems in person ReID: first, it is difficult to locate the target person because the person is frequently partially occluded in crowed scenes; second, it is difficult to match the target person due to the similarity of the target person and other pedestrian features; third, the problem of model performance degradation caused by the large style discrepancies across domain/datasets. These three problems greatly limit the application of person ReID in real scenes. To solve these problems, we proposed a person ReID method based on effective features and self-optimized pseudo-label. Firstly, we designed a feature aggregation module which combines mask channel and pose channel to accurately extract the global saliency features, so as to solve the occlusion problem; secondly, we designed a head-shoulder feature auxiliary module to enhance the feature representation of the head-shoulder, so as to solve the problem of similarity between the target person and other pedestrian features; finally, we designed a self-optimized pseudo-label training module to improves the generalization ability of the model, so as to solve the problem of different styles in the cross-domain environment. Extensive contrast experiments with the state-of-the-art methods on multiple person re-ID datasets show that our method leads to significant improvement, which prove the effectiveness of our method.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.