Traditional machine learning methods struggle with plant pest and disease image recognition, particularly when dealing with small sample sizes, indistinct features, and numerous categories. This paper proposes an improved ResNet34 model (ESA-ResNet34) for crop pest and disease detection. The model employs ResNet34 as its backbone and introduces an efficient spatial attention mechanism (effective spatial attention, ESA) to focus on key regions of the images. By replacing the standard convolutions in ResNet34 with depthwise separable convolutions, the model reduces its parameter count by 85.37% and its computational load by 84.51%. Additionally, Dropout is used to mitigate overfitting, and data augmentation techniques such as center cropping and horizontal flipping are employed to enhance the model’s robustness. The experimental results show that the improved algorithm achieves an accuracy, precision, and F1 score of 87.09%, 87.14%, and 86.91%, respectively, outperforming several benchmark models (including AlexNet, VGG16, MobileNet, DenseNet, and various ResNet variants). These findings demonstrate that the proposed ESA-ResNet34 model significantly enhances crop pest and disease detection.

Ensuring the structural continuity and completeness of road networks in high-resolution remote sensing imagery remains a major challenge for current deep learning methods, especially under conditions of occlusion caused by vegetation, buildings, or shadows. To address this, we propose a novel post-processing enhancement framework that improves the connectivity and accuracy of initial road extraction results produced by any segmentation model. The method employs a dual-stream encoder architecture, which jointly processes RGB images and preliminary road masks to obtain complementary spatial and semantic information. A core component is the MAF (Multi-Modal Attention Fusion) module, designed to capture fine-grained, long-range, and cross-scale dependencies between image and mask features. This fusion leads to the restoration of fragmented road segments, the suppression of noise, and overall improvement in road completeness. Experiments on benchmark datasets (DeepGlobe and Massachusetts) demonstrate substantial gains in precision, recall, F1-score, and mIoU, confirming the framework’s effectiveness and generalization ability in real-world scenarios.

Learning with noisy labels (LNL) remains challenging, especially when the identification of clean samples relies heavily on the predictions of the model being trained. In such cases, early-stage selection errors may be reinforced during iterative optimization, leading to unreliable supervision. To alleviate this issue, a two-stage framework, termed CANNE, is proposed by combining Contrastive Language–Image Pre-training (CLIP)-based conservative offline cleaning with Adaptive Nearest Neighbors and eigenvector-based sample selection (ANNE)-based online refinement. Specifically, a high-confidence clean seed set is first constructed using two complementary probability sources derived from frozen CLIP representations and reliability criteria, including class-wise loss modeling and prediction consistency. This seed set is then used as a set of reliable anchors during the subsequent ANNE training process, where online feature- and neighborhood-based refinement further recovers and adjusts sample partitions. In this way, CANNE uses external vision–language priors to provide conservative and persistent guidance while preserving the adaptive recovery ability of online noisy-label learning. Experimental results on CIFAR-10, CIFAR-100, Animal-10N, and Mini-WebVision, together with additional evaluation under open-set noise, show that the proposed method achieves competitive performance across diverse noisy-label settings. In particular, CANNE achieves 96.6% and 96.3% best accuracies on CIFAR-10 under 80% and 90% symmetric noise, respectively, and 81.0% and 79.0% on CIFAR-100 under 20% and 50% symmetric noise. Additional repeated-run, threshold-sensitivity, and runtime analyses further indicate that the CLIP-based seed set provides stable guidance with only moderate computational overhead.

Deep neural networks (DNNs), due to their over-parameterization and memorization capacity, are prone to fitting incorrect labels in training data, which can substantially degrade their performance. Consequently, LNL has attracted increasing attention. Popular LNL methods select potentially clean samples for reliable supervision while applying robust techniques to the remaining samples to mitigate the adverse effects of label noise. However, when sample selection relies solely on the continuously updated classifier, the classifier may gradually memorize incorrect labels during training, and the resulting biased predictions can further affect subsequent sample selection, leading to a self-reinforcing confirmation bias. To address this issue, we propose CAMN, a CLIP-guided adaptive sample selection with machine unlearning and negative learning strategies for noisy labels learning. Specifically, CAMN introduces a pretrained vision language model CLIP as an external semantic prior independent of the task-specific classifier to perform preliminary noise-aware sample selection. The resulting selections are then integrated with an adaptive and balanced sample selection strategy to provide more reliable samples for robust learning. CLIP provides an independent external semantic prior for reducing classifier-induced bias, while the selection results from the trained DNN complements CLIP-based zero-shot selection by reducing over-reliance on its predictions and promoting more balanced sample selection. After selection, CAMN further incorporates a machine unlearning mechanism and a negative learning strategy as regularizers, reducing the persistent influence of noisy labels on DNN and further enhancing its robustness. Experiments demonstrate the effectiveness and robustness of CAMN across diverse noise settings, including synthetic and Web-crawled noisy datasets.

Recently, unsupervised deep learning based on the Retinex model for low‐light image enhancement (LLIE) has made significant progress. However, most existing methods typically assume gradual illumination changes and use single‐path network to first estimate illumination for image decomposition, then correct it for low‐light enhancement. This not only makes it difficult to effectively address low‐light noise and rapid illumination changes but also overlooks the potential amplification of original reflectance contrast, which limits their performance. To address these challenges, this work proposes an efficient algorithm called MDSC for LLIE, which comprises three functional modules: detail‐aware image smoothing (DAS), multipath image decomposition (MPD), and reflectance synergistic correction (RSC). Concretely, the input low‐light image is first smoothed with DAS to suppress noise while preserving details. Next, the smoothed image is processed by MPD to perform multipath directed decomposition and fusion, mitigating rapid variations in illumination and producing more accurate illumination and reflectance components. Finally, RSC jointly calibrates the illumination and reflectance components and reilluminates the scene to counteract potential amplification of reflectance contrast and produce the final enhanced result. Experimental results on challenging datasets demonstrate the effectiveness of the proposed MDSC, with both quantitative and qualitative outcomes significantly outperforming most state‐of‐the‐art approaches.