To address issues such as low accuracy in vehicle and pedestrian detection and slow inference speed, an improved version of the FCOS object detection algorithm is proposed. The algorithm replaces the original backbone with ResNet18 and incorporates deformable convolution to enhance its ability to capture shape features of the targets, thereby improving feature extraction and model inference speed. A coarse localization box mechanism is added to the algorithm's localization detection head branch to improve the model's classification accuracy and detection box recall. A new background prediction branch is introduced to enhance the model's detection capability for positive samples by predicting the probability of differences with the environment. Additionally, an ATSS (Adaptive Training Sample Selection) label assignment scheme and GFocal Loss function are employed to replace the original label assignment scheme and loss function, mitigating the impact of imbalances between positive and negative samples in the dataset. The experimental results show that the improved algorithm achieves 94.7% on the KITTI traffic target data set, which is 22.2% higher than the original FCOS model, improves the inference speed to 74 pieces per second, and has better effect compared with some mainstream target detection algorithms.
The recognition effect of news text sequence data is strongly related to the importance of each word and the dependency relationship between them. Although the capsule network can learn the correlation information between news text as a whole and local, it lacks the attention to the key words in the text and ignores the distant dependencies in the text. To remedy the above shortcomings, this paper proposes a news text classification model which is based on multi-head attention and parallel capsule networks, using a multi-head attention layer for feature extraction and then a parallel capsule network module as the classification layer. The model can retrieve wealthier text details. Experimental results demonstrate that the proposed model of this paper works better than the mainstream capsule network based text classification models in both single-label and multi-label classification tasks of news texts.
Aiming at the sparse features of Chinese short texts, a multi-granularity information-enhanced text emotion classification model is proposed, which combines convolutional neural network and attention mechanism. First, fine-tune BERT model for specific tasks to obtain different granularity information contained in different layers, and express the text at sentence level as character-level vector. Then, using the feature of CNN with local semantic feature extraction, the multi-granularity local semantics of each layer of BERT is extracted with different scale convolutional kernels to obtain the multi-granularity local semantic representation under different layers. Considering that CLS at different levels of BERT contains global information at different levels of the text, it integrates and interacts with multi-granularity local semantic representation through attention mechanism to obtain multi-granularity deep semantic information. Finally, the final global semantic representation of BERT is fused with it to obtain the final text representation, which is input to the classifier for classification. The experimental comparison between this model and several models shows that the indexes of this model on two real data sets are obviously improved compared with the comparison model, which proves the effectiveness of this model.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.