Yunzhi Wu

Accurate segmentation of pear leaf diseases is paramount for enhancing diagnostic precision and optimizing agricultural disease management. However, variations in disease color, texture, and morphology, coupled with changes in lighting conditions and gradual disease progression, pose significant challenges. To address these issues, we propose EBMA-Net, an edge-aware multi-scale network. EBMA-Net introduces a Multi-Dimensional Joint Attention Module (MDJA) that leverages atrous convolutions to capture lesion information at different scales, enhancing the model’s receptive field and multi-scale processing capabilities. An Edge Feature Extraction Branch (EFFB) is also designed to extract and integrate edge features, guiding the network’s focus toward edge information and reducing information redundancy. Experiments on a self-constructed pear leaf disease dataset demonstrate that EBMA-Net achieves a Mean Intersection over Union (MIoU) of 86.25%, Mean Pixel Accuracy (MPA) of 91.68%, and Dice coefficient of 92.43%, significantly outperforming comparison models. These results highlight EBMA-Net’s effectiveness in precise pear leaf disease segmentation under complex conditions.

Chinese Optical Character Recognition (OCR) technology is essential for digital transformation in Chinese regions, enabling automated document processing across various applications. However, Chinese OCR systems struggle with visually similar characters, where subtle stroke differences lead to systematic recognition errors that limit practical deployment accuracy. This study develops FLIP (Feedback Learning-based Intelligent Plugin), a lightweight post-processing plugin designed to improve Chinese OCR accuracy across different systems without external dependencies. The plugin operates through three core components as follows: UTF-8 encoding-based output parsing that converts OCR results into mathematical representations, error correction using information entropy and weighted similarity measures to identify and fix character-level errors, and adaptive feedback learning that optimizes parameters through user interactions. The approach functions entirely through mathematical calculations at the character encoding level, ensuring universal compatibility with existing OCR systems while effectively handling complex Chinese character similarities. The plugin’s modular design enables seamless integration without requiring modifications to existing OCR algorithms, while its feedback mechanism adapts to domain-specific terminology and user preferences. Experimental evaluation on 10,000 Chinese document images using four state-of-the-art OCR models demonstrates consistent improvements across all tested systems, with precision gains ranging from 1.17% to 10.37% and overall Chinese character recognition accuracy exceeding 98%. The best performing model achieved 99.42% precision, with ablation studies confirming that feedback learning contributes additional improvements from 0.45% to 4.66% across different OCR architectures.

The accurate pest control of pear tree diseases is an urgent need for the realization of smart agriculture, with one of the key challenges being the precise segmentation of pear leaf diseases. However, existing methods show poor segmentation performance due to issues such as the small size of certain pear leaf disease areas, blurred edge details, and background noise interference. To address these problems, this paper proposes an improved U-Net architecture, FFAE-UNet, for the segmentation of pear leaf diseases. Specifically, two innovative modules are introduced in FFAE-UNet: the Attention Guidance Module (AGM) and the Feature Enhancement Supplementation Module (FESM). The AGM module effectively suppresses background noise interference by reconstructing features and accurately capturing spatial and channel relationships, while the FESM module enhances the model’s responsiveness to disease features at different scales through channel aggregation and feature supplementation mechanisms. Experimental results show that FFAE-UNet achieves 86.60%, 92.58%, and 91.85% in MIoU, Dice coefficient, and MPA evaluation metrics, respectively, significantly outperforming current mainstream methods. FFAE-UNet can assist farmers and agricultural experts in more effectively evaluating and managing diseases, thereby enabling precise disease control and management.

In agricultural production, pest problems are inevitable. In recent years, China has lost up to 40 million tons of grain annually due to various pests and diseases. In nature, with over one million known insect species, traditional manual identification is costly, inefficient and difficult to accurately distinguish all pests. With the development of computer technology, vision-based unimodal recognition models have been proposed. Due to the complex background of pest habitats and the diverse morphologies of pests, unimodal recognition models often lead to issues such as insufficient identification accuracy and low data utilization rates. As smart agriculture advances, multimodal data continues to emerge. Therefore, current research is focused on cross-modal data fusion to compensate for the deficiencies of unimodal recognition. However, research at this stage tends to overlook the importance of cross-modal hierarchical features and issues such as redundancy in textual features. In order to solve the above problems, this paper proposes a multiscale intermodal feature fusion model (ITFNet), which fuses multiscale visual and textual features for intermodal attention learning, enabling the model to effectively utilize the complementary nature of multimodal data. Specifically, pest text descriptions are designed to aggregate manual template features in feature extraction via language model (LM) to obtain text features; meanwhile, pest image data are used to obtain visual features at different levels via a five-layer cascaded downsampling module. To enhance the model's attention to key regions, the extracted visual features are fused with textual features after being processed by an attention mechanism. Compared to the traditional unimodal YOLOv4 model, the model proposed in this paper achieves an 18.02% improvement in the mAP (mean Average Precision) metric. In addition, in order to prove the uniqueness of high-level features in cross-modality, this study also conducts ablation experiments on different levels of visual and textual features, and the results show that the model's mAP value has a 2.38% improvement compared to the full fusion after the exclusion of high-level visual feature fusion, which further proves the importance of the selective fusion strategy. In summary, the multiscale cross-modal feature fusion model proposed in this paper has practical application value for advancing the development of multimodal modeling in the field of agriculture, enhancing the depth and breadth of agricultural pest control research, and even improving the efficiency of agricultural production.

In agricultural production, pest problems are inevitable. In recent years, China has suffered annual losses of up to 40 million tons of grain due to various pests and diseases. There are over one million known insect species in the natural world, exhibiting complex and diverse morphologies, making manual identification costly. With the advancement of deep learning technologies, methods relying solely on image data for crop pest identification have achieved some success. However, these methods heavily depend on numerous high-quality annotated images. Many existing approaches overlook the role and value of other modalities beyond images and rely solely on low-level image features for recognition. They fail to fully exploit semantic correlations between multimodal data, leading to poor reliability and interpretability of the recognition results. Moreover, in real-world scenarios, pest habitats are complex and diverse, with varying scales, imposing high demands on the practical application of models. To address these challenges, this paper proposes a Multi-Scale Cross-Modal Feature Fusion Model (ITFNet-API) utilizing multi-scale visual features and textual features for cross-modal attention learning. Super-resolution reconstruction techniques are introduced to restore high-frequency information in low-quality images. To avoid altering low-frequency information during the reconstruction process, the reconstructed high-quality images are combined with the original images and fed into the model for feature learning to enrich pest features. The proposed Image and Text Fusion Module (ITFM) performs multi-scale cross-modal information fusion and an Inverse Transpose Convolution (ITC) network module is introduced to restore multi-scale visual features. The obtained visual features are input into the neck and head networks of the YOLOv4 (You Only Look Once version 4) model to obtain pest category and location information in each image. Finally, to verify the model’s practical application capability, this paper introduces an Arbitrary Combination Image Enhancement (ACIE) data augmentation method which generates a batch of complex multi-target, multi-scale pest images in challenging scenarios for data augmentation and model training. We conducted extensive comparative experiments based on the IP102 dataset to validate the effectiveness of the proposed method. The experimental results indicate that our method achieved an accuracy rate of 82.15% in pest recognition, outperforming other advanced methods significantly with an AP50 of 92.18%. Moreover, in multi-target recognition,the accuracy rate reached 45.24% with an AP50 of 43.04%.

The accurate prevention and control of pear tree diseases, especially the precise segmentation of leaf diseases, poses a serious challenge to fruit farmers globally. Given the possibility of disease areas being minute with ambiguous boundaries, accurate segmentation becomes difficult. In this study, we propose a pear leaf disease segmentation model named MFBP-UNet. It is based on the UNet network architecture and integrates a Multi-scale Feature Extraction (MFE) module and a Tokenized Multilayer Perceptron (BATok-MLP) module with dynamic sparse attention. The MFE enhances the extraction of detail and semantic features, while the BATok-MLP successfully fuses regional and global attention, striking an effective balance in the extraction capabilities of both global and local information. Additionally, we pioneered the use of a diffusion model for data augmentation. By integrating and analyzing different augmentation methods, we further improved the model’s training accuracy and robustness. Experimental results reveal that, compared to other segmentation networks, MFBP-UNet shows a significant improvement across all performance metrics. Specifically, MFBP-UNet achieves scores of 86.15%, 93.53%, 90.89%, and 0.922 on MIoU, MP, MPA, and Dice metrics, marking respective improvements of 5.75%, 5.79%, 1.08%, and 0.074 over the UNet model. These results demonstrate the MFBP-UNet model’s superior performance and generalization capabilities in pear leaf disease segmentation and its inherent potential to address analogous challenges in natural environment segmentation tasks.

Picture response is a crucial aspect in dialogue systems; however, existing systems predominantly focus on text information, often neglecting the use of image data or relying on large-scale language models, which leads to high training costs and slower generation speeds. To address these challenges, this paper introduces the MDSAGM (Selective Attention and Gating Mechanism)model, which comprises a Transformer-based Text Dialogue Response Generator, Text-to-Image Generation, and SAGM (Selective Attention and Gating Mechanism). The model is core lies in combining Selective Attention and Gating Mechanism to enhance generalization and improve model accuracy. The study further explores the contributions of image and text information in multimodal fusion. Compared to other large-scale training models, the proposed model demonstrates computational efficiency , fewer parameters, and shorter response times, making it more lightweight. This dialogue model facilitates the separation of multimodal dialogue parameters from the overall model, enabling better parameter fitting through pre-training with abundant plain text and text-image data. Extensive experiments show that the proposed method achieves promising results in both automatic and manual evaluations, generating information-rich text and image responses with higher accuracy while maintaining a lightweight structure.

Image datasets acquired from orchards are commonly characterized by intricate backgrounds and an imbalanced distribution of disease categories, resulting in suboptimal recognition outcomes when attempting to identify apple leaf diseases. In this regard, we propose a novel apple leaf disease recognition model, named RFCA ResNet, equipped with a dual attention mechanism and multi-scale feature extraction capacity, to more effectively tackle these issues. The dual attention mechanism incorporated into RFCA ResNet is a potent tool for mitigating the detrimental effects of complex backdrops on recognition outcomes. Additionally, by utilizing the class balance technique in conjunction with focal loss, the adverse effects of an unbalanced dataset on classification accuracy can be effectively minimized. The RFB module enables us to expand the receptive field and achieve multi-scale feature extraction, both of which are critical for the superior performance of RFCA ResNet. Experimental results demonstrate that RFCA ResNet significantly outperforms the standard CNN network model, exhibiting marked improvements of 89.61%, 56.66%, 72.76%, and 58.77% in terms of accuracy rate, precision rate, recall rate, and F1 score, respectively. It is better than other approaches, performs well in generalization, and has some theoretical relevance and practical value.

Insect pests are a major element influencing agricultural production. According to the Food and Agriculture Organization (FAO), an estimated 20–40% of pest damage occurs each year, which reduces global production and becomes a major challenge to crop production. These insect pests cause sooty mold disease by sucking the sap from the crop’s organs, especially leaves, fruits, stems, and roots. To control these pests, pesticides are frequently used because they are fast-acting and scalable. Due to environmental pollution and health awareness, less use of pesticides is recommended. One of the salient approaches could be to reduce the wide use of pesticides by spraying on demand. To perform spot spraying, the location of the pest must first be determined. Therefore, the growing population and increasing food demand emphasize the development of novel methods and systems for agricultural production to address environmental concerns and ensure efficiency and sustainability. To accurately identify these insect pests at an early stage, insect pest detection and classification have recently become in high demand. Thus, this study aims to develop an object recognition system for the detection of crops damaging insect pests and their classification. The current work proposes an automatic system in the form of a smartphone IP- camera to detect insect pests from digital images/videos to reduce farmers’ reliance on pesticides. The proposed approach is based on YOLO object detection architectures including YOLOv5 (n, s, m, l, and x), YOLOv3, YOLO-Lite, and YOLOR. For this purpose, we collected 7046 images in the wild under different illumination and background conditions to train the underlying object detection approaches. We trained and test the object recognition system with different parameters from scratch. The eight models are compared and analyzed. The experimental results show that the average precision (AP@0.5) of the eight models including YOLO-Lite, YOLOv3, YOLOR, and YOLOv5 with five different scales (n, s, m, l, and x) reach 51.7%, 97.6%, 96.80%, 83.85%, 94.61%, 97.18%, 97.04%, and 98.3% respectively. The larger the model, the higher the average accuracy of the detection validation results. We observed that the YOLOv5x model is fully functional and can correctly identify the twenty-three species of insect pests at 40.5 milliseconds (ms). The developed model YOLOv5x performs the state-of-the-art model with an average precision value of (mAP@0.5) 98.3%, (mAP@0.5:0.95) value of 79.8%, precision of 94.5% and a recall of 97.8%, and F1-score with 96% on our IP-23 dataset. The results show that the system works efficiently and was able to correctly detect and identify insect pests, which can be employed for realistic application while farming.