Xichen Yang

Recently, CLIP has shown strong potential in no-reference image quality assessment (NR-IQA) tasks due to its powerful cross-modal alignment capability. However, existing methods mainly treat CLIP as a feature extractor and train models using the final encoded features, while the application of interactions within the intermediate layers of the image and text encoders is insufficient. Moreover, the methods focus on visual distortions without sufficiently employing the content characteristics. To address these issues, the paper proposes a NR-IQA method based on visual-textual fusion via cross-attention in dual-stream CLIP. Specifically, a dual-stream CLIP architecture is constructed, with one stream extracting content-aware features, and the other extracting quality-aware features. Meanwhile, a cross-attention module is designed to enable deep fusion between visual and textual features, yielding more comprehensive quality representations. In addition, a joint loss function combining quality regression loss and cross-branch image-text consistency constraints is proposed, allowing for adaptive balancing of branch weights while preserving the unique characteristics of each stream. Experimental results on public datasets demonstrate that the proposed method achieves competitive performance compared with existing methods, confirming its effectiveness across diverse scenarios. The codes and models are available at https://github.com/dart-into/DCLIP-CA.

Chrysanthemum is a traditional Chinese medicinal herb that contains significant medicinal and economic value. However, the medicinal and economic value of chrysanthemum differs depending on its regions and types. Therefore, it is valuable to classify chrysanthemum accurately. Traditional classification methods are costly, time-consuming, and mainly rely on manual processes, chemical testing, or genetic analysis. In light of these challenges, this paper proposes a Chrysanthemum Classification via Color Space Fusion Transformer, which is both cost-effective and capable of real-time processing. First, the chrysanthemum images in the RGB color space are converted to the LAB color space. Second, a multi-path network is designed to independently extract color space features from both the RGB and LAB color spaces, followed by their integration through an inter-path fusion module. Finally, the Transformer module further analyzes the semantic characteristics of these extracted color space features. Experimental results indicate that the proposed method achieves superior accuracy and stability compared to existing classification methods, with a classification accuracy of 96.16%. This method provides an efficient and practical solution for chrysanthemum origin traceability.

Citri Reticulatae Pericarpium (CRP), the dried peel of citrus fruits, holds notable dietary and medicinal value. Its quality and price largely depend on origin and aging. Lower-grade CRP is often adulterated to imitate premium products, making accurate authentication of region and vintage essential for quality assurance and fair market valuation. Existing methods for vintage classification are limited due to complex equipment and high operational costs, restricting their scalability in practical applications. To address these issues, a convenient method for the accurate identification of Citri Reticulatae Pericarpium using image and multi-stream is proposed. The method comprises three main stages. Firstly, an object detection network with bounding box refinement localizes exocarp and albedo regions from whole CRP images. Secondly, a three-stream feature extractor processes the whole images along with exocarp and albedo patches to capture complementary visual details. A channel-level feature interaction module further enhances robustness through cross-region feature integration. Thirdly, a meta-learning module enables rapid adaptation to images captured under varying conditions by different consumer-grade devices. Experimental results demonstrate that the proposed method achieves an accuracy of 95.5% on iPhone-captured images. In addition, for images captured by different devices, the proposed method achieves a relative accuracy improvement of more than 34% over the direct transfer method, mainly owing to the meta-learning adaptation to different devices.

Most existing image quality assessment methods need to be retrained when dealing with a new type of task. This approach wastes computing resources and time. Therefore, these methods fail to suit the application scenarios that require processing of multi-type image quality assessment tasks. In the human visual system, the eyes of human tend to pay varying degrees of attention to different regions. Inspired by this system, this paper proposes a multi-type image quality assessment method based on multi-region deep feature fusion under meta-learning (MMQA). First, we utilize the differences in the structural information to screen out salient and non-salient regions. Second, a deep multi-stream network is designed to comprehensively consider and fuse different features related to the quality in salient regions, non-salient regions and the entire image. Third, meta-learning is applied to quickly learn and update the parameters of the model when facing new types of images. By summarizing the prior knowledge in the training of one type of task, the model can be quickly fine-tuned for other types of images. The experimental results demonstrate that the proposed method has advantages over the existing methods in generalization and robustness. Furthermore, the proposed method can adapt well to different distortion types and different image types quickly and accurately.