Specialized pest control for agriculture is a high-priority agricultural issue. There are multiple categories of tiny pests, which pose significant challenges to monitoring. Previous work mainly relied on manual monitoring of pests, which was labor-intensive and time-consuming. Recently, deep-learning-based pest detection methods have achieved remarkable improvements and can be used for automatic pest monitoring. However, there are two main obstacles in the task of pest detection. (1) Small pests often go undetected because much information is lost during the network training process. (2) The highly similar physical appearances of some categories of pests make it difficult to distinguish the specific categories for networks. To alleviate the above problems, we proposed the multi-category pest detection network (MCPD-net), which includes a multiscale feature pyramid network (MFPN) and a novel adaptive feature region proposal network (AFRPN). MFPN can fuse the pest information in multiscale features, which significantly improves detection accuracy. AFRPN solves the problem of anchor and feature misalignment during RPN iterating, especially for small pest objects. In extensive experiments on the multi-category pests dataset 2021 (MPD2021), the proposed method achieved 67.3% mean average precision (mAP) and 89.3% average recall (AR), outperforming other deep learning-based models.

It is well recognized that aphid infestation severely reduces crop yield and further leads to significant economic loss. Therefore, accurately and efficiently detecting aphids is of vital importance in pest management. However, most existing detection methods suffer from unsatisfactory performance without fully considering the aphid characteristics, including tiny size, dense distribution, and multi-viewpoint data quality. In addition, existing clustered tiny-sized pest detection methods improve performance at the cost of time and do not meet the real-time requirements. To address the aforementioned issues, we propose a robust aphid detection method with two customized core designs: a Transformer feature pyramid network (T-FPN) and a multi-resolution training method (MTM). To be specific, the T-FPN is employed to improve the feature extraction capability by a feature-wise Transformer module (FTM) and a channel-wise feature recalibration module (CFRM), while the MTM aims at purifying the performance and lifting the efficiency simultaneously with a coarse-to-fine training pattern. To fully demonstrate the validity of our methods, abundant experiments are conducted on a densely clustered tiny pest dataset. Our method can achieve an average recall of 46.1% and an average precision of 74.2%, which outperforms other state-of-the-art methods, including ATSS, Cascade R-CNN, FCOS, FoveaBox, and CRA-Net. The efficiency comparison shows that our method can achieve the fastest training speed and obtain 0.045 s per image testing time, meeting the real-time detection. In general, our TD-Det can accurately and efficiently detect in-field aphids and lays a solid foundation for automated aphid detection and ranking.

Multimodal contrastive learning has achieved significant performance advantages in self-supervised skeleton-based action recognition. Previous methods are limited by modality imbalance, which reduces alignment accuracy and makes it difficult to combine important spatial–temporal frequency patterns, leading to confusion between modalities and weaker feature representations. To overcome these problems, we explore intra-modality feature-wise self-similarity and inter-modality instance-wise cross-consistency, and discover two inherent correlations that benefit recognition: (i) Global Perspective expresses how action semantics carry a broad and high-level understanding, which supports the use of globally discriminative feature representations. (ii) Focus Adaptation refers to the role of the frequency spectrum in guiding attention toward key joints by emphasizing compact and salient signal patterns. Building upon these insights, we propose a novel language–skeleton contrastive learning framework comprising two key components: (a) Feature Modulation, which constructs a skeleton–language action conceptual domain to minimize the expected information gain between vision and language modalities. (b) Frequency Feature Learning, which introduces a Frequency-domain Spatial–Temporal block (FreST) that focuses on sparse key human joints in the frequency domain with compact signal energy. Extensive experiments demonstrate the effectiveness of our method achieves remarkable action recognition performance on widely used benchmark datasets, including NTU RGB+D 60 and NTU RGB+D 120. Especially on the challenging PKU-MMD dataset, MICA has achieved at least a 4.6% improvement over classical methods such as CrosSCLR and AimCLR, effectively demonstrating its ability to capture internal and contextual attention information.

Joint training across heterogeneous paired underwater image enhancement (UIE) datasets broadens the range of supervised training conditions but requires a shared model to learn how the mapping from an observation to its reference varies across source domains while preserving color fidelity and spatial detail. We study a balanced source-known joint-training setting, in which one shared model is trained across multiple paired datasets and DCSM-Net receives the dataset-of-origin index during both training and evaluation. We propose a Domain-Conditioned Spatial–Frequency Modulation Network (DCSM-Net) for this setting. DCSM-Net integrates three lightweight operations within a hybrid encoder–decoder. A Spatial-Guided (SG) block performs high-resolution local refinement and applies channel-asymmetric latent feature modulation initialized with a wavelength-motivated prior. A Domain Token Module (DTM) combines an image-context descriptor with a learnable source-domain embedding to generate a conditioning token. Conditional Wavelet Prompt Blocks (WPBs) use this token to modulate skip features, applying channel-wise affine transformation to the low-frequency LL component and residual prompt routing to the directional LH, HL, and HH components. Under joint training on UIEB, LSUI, and EUVP, DCSM-Net improves the average PSNR of the same backbone from 25.51 to 26.75 dB with only 0.13 M additional parameters. Under this source-known protocol, it achieves 25.13 dB on UIEB-Test90, 31.69 dB on LSUI-Test150, and 23.42 dB on EUVP-Test100, obtaining the best PSNR, SSIM, and LPIPS among the compared methods. In an additional DCSM-Net analysis, we assess seed sensitivity across seeds 7, 17, and 27 and examine a source-agnostic mean-embedding fallback; the main comparisons otherwise use seed 7. Hierarchical ablations show gains from high-resolution spatial refinement, statistics-conditioned wavelet processing, and source-domain token conditioning.

AI-generated text detectors often achieve high in-domain accuracy but remain sensitive to changes in domain, generator, and writing style. This limitation partly arises because conventional detectors rely predominantly on latent language-model representations and fixed decision boundaries, without explicitly characterizing sample-level structural and semantic variation. We introduce S²E-AL, an end-to-end score-aware structure–semantic entropy adversarial learning framework for AI-generated text detection. S²E-AL encodes local dependency, global contextual, and semantic entropy features into a structure–semantic representation. A gradient-reversal source classifier discourages this representation from retaining source-specific information, while a score-aware threshold estimator conditions on both the entropy representation and a detached base-detector logit to produce a sample-specific logit adjustment. The adjusted logit is used to estimate the final probability that a text is AI-generated. Across HC3, HC3_Plus, LLMTrace, RAID, and XSum, the score-aware model achieved a macro Accuracy of 90.90% and a macro F1-score of 91.42%. Its performance remained lower on XSum, indicating that substantial domain and stylistic shifts remain challenging. These results suggest that jointly modeling structure–semantic variation and detector confidence can improve the overall performance–stability trade-off of AI-generated text detection across heterogeneous test distributions.