Earf-YOLO: An Efficient Attention Receptive Field Model for Recognizing Symbols of Zhuang Minority Patterns

Abstract

As for recognizing Zhuang minority pattern symbols, current recognition models often cause high computational overhead and low accuracy since Zhuang minority pattern symbols have large feature vectors and some complex features. In this paper, we present the efficient attention receptive field you only look once (Earf-YOLO), a new scheme to address those problems. Firstly, a global-local-transformer (GLocalT) structure is proposed, through which other control systems are introduced into the axial self-attention module, and global-local training strategies are also designed. The structure can use other control systems to compensate for the lost feature information along the height, width, and channel axes. The global-local training strategy can encode long-term dependencies between features and reduce local information loss, fully illustrating that the structure has high feature expression ability. Besides, strength receptive field block (SRFB) is suggested to use the dilated convolution to control the receptive field’s eccentricity and enrich the feature information of the receptive field during its training. With more branches, it can better extract multiscale features, enrich the feature space of the convolution block, and reparametrize multibranch during prediction to fuse them into the main branch, all of which contribute to the improvement of the model performance. Finally, some advanced training techniques are adopted to enhance the detection effect further. In the end, comparative experiments are conducted on the datasets of Zhuang pattern symbols and PASCAL VOC, whose results indicate that the AP and FPS of the suggested model reach their highest values, manifesting its high efficiency.

Transformer-based comparative multi-view illegal transaction detection

“…The Transformer is able to process the input data in a parallel manner and effectively solve the long-time dependency problem, significantly reducing training and prediction time. The Transformer has demonstrated significant benefits in data processing in a number of areas such as rain removal [34], target recognition [35], etc. The Transformer model consists of a Position embedding module, a multi-headed attention mechanism and a feed-forward network.…”

Improving image super-resolution with structured knowledge distillation-based multimodal denoising diffusion probabilistic model

“…Deep learning-based methods have excelled in feature extraction, achieving remarkable successes in various traditional tasks such as fault detection, [1][2][3] deraining, 4 medical segmentation, 5 and object detection. 6 These methods are particularly suited for establishing mappings between HR and LR images, especially in handling the empirical distribution of complex images. In the context of SISR tasks, deep learning-based methods can be categorized into several types, primarily including generative adversarial networks (GANs), 7 variational autoencoders (VAEs), 8,9 and normalizing flows.…”

“…Deep learning-based methods have excelled in feature extraction, achieving remarkable successes in various traditional tasks such as fault detection, 1 – 3 deraining, 4 medical segmentation, 5 and object detection 6 . These methods are particularly suited for establishing mappings between HR and LR images, especially in handling the empirical distribution of complex images.…”