Zhiqiang Zhao

Reliable manual annotations for Vietnamese logistics calls are scarce, and telephone channel degradation compounds errors in business-critical entities. Existing adaptation methods typically optimize aggregate transcription accuracy, use uniform token losses, or apply a fixed training schedule; few control parameter updates, entity-level supervision, and sample order together when labels are noisy. We address this problem by combining low-rank adaptation (LoRA), token-level entity-weighted cross-entropy, and validation feedback curriculum learning for Whisper-large-v3. A joint acoustic–semantic score ranks samples by signal quality, sequence length, entity density, and linguistic rarity. The curriculum boundary expands only after the validation loss improves. On 20,317 in-domain training utterances, the complete system achieves a word error rate (WER) of 0.0735 and a mean exact match entity accuracy of 92.1%. Compared with vanilla LoRA, it reduces WER by 4.92% and raises entity accuracy from 85.3% to 92.1%. Ablation results show that curriculum pacing and entity weighting each contribute to the final performance.

With the rapid development of information technology in education, teaching resources have increased dramatically and the teaching environment has improved. However, resource overload and learning disorientation are challenges faced by learners. This study proposes a personalized learning path generation model based on neurocognitive collaborative filtering technology. By analyzing learners' personalized static parameters and dynamic learning behavior data, explicit and implicit weak knowledge points are identified, and optimized learning paths are generated. The empirical results show that the method significantly improves learners' learning outcomes and satisfaction over traditional methods. This study promotes the development of personalized learning design and effective learning support systems with important implications for educational practice.

Crowd counting is an important task that serves as a preprocessing step in many applications. Despite obvious improvement reported by various convolutional-neural-network-based approaches, they only focus on the role of deep feature maps while neglecting the importance of shallow features for crowd counting. In order to surmount this issue, a dilated convolutional-neural-network-based cross-level contextual information extraction network is proposed in this work, which is abbreviated as CL-DCNN. Specifically, a dilated contextual module (DCM) is constructed by importing cross-level connection between different feature maps. It can effectively integrate contextual information while conserving the local details of crowd scenes. Extensive experiments show that the proposed approach outperforms state-of-the-art approaches using five public datasets, i.e., ShanghaiTech part A, ShanghaiTech part B, Mall, UCF_CC_50 and UCF-QNRF, achieving MAE 52.6, 8.1, 1.55, 181.8, and 96.4, respectively.

Change detection from synthetic aperture radar (SAR) images is of great significance for natural environmental protection and human societal activity, which can be regarded as the process of assigning a class label (changed or unchanged) to each of the image pixels. This paper presents a novel classification technique to address the SAR change-detection task that employs a generalized Gamma deep belief network (gΓ-DBN) to learn features from difference images. We aim to develop a robust change detection method that can adapt to different types of scenarios for bitemporal co-registered Yellow River SAR image data set. This data set characterized by different looks, which means that the two images are affected by different levels of speckle. Widely used probability distributions offer limited accuracy for describing the opposite class pixels of difference images, making change detection entail greater difficulties. To address the issue, first, a gΓ-DBN can be constructed to extract the hierarchical features from raw data and fit the distribution of the difference images by means of a generalized Gamma distribution. Next, we propose learning the stacked spatial and temporal information extracted from various difference images by the gΓ-DBN. Consequently, a joint high-level representation can be effectively learned for the final change map. The visual and quantitative analysis results obtained on the Yellow River SAR image data set demonstrate the effectiveness and robustness of the proposed method.

With the development of synthetic aperture radar (SAR) imaging techniques, many high-quality images are acquired for exactly earth observation. As a result of heavy speckle, it is difficult to extract effective representation of different scene images to distinguish them accurately. In this paper, a metric-learning based convolutional neural network (CNN) is proposed to characterize SAR images. Specifically, a regularization term is proposed, which is constrained by the Mahalanobis distance. The regularization term learns a new metric space based on CNN, where the similar images should be close and those of dissimilar images should be separated. Comparing with the other approaches, experimental results of the proposed approach illustrate a superior classification for different scenes.

With the increased number of synthetic aperture radar (SAR) systems that have been launched in recent years, many high-quality images have been produced for precise earth observation. The number of available images requires new feature extraction techniques to capture enough discriminant information from high-resolution SAR images. To effectively characterize different land covers, a multiscale local Fisher pattern (MLFP) is proposed to extract both statistical and spatial properties in this work. First, the Fisher vector (FV) of each pixel is calculated after the statistical distribution of high-resolution SAR intensities is modeled via a Gamma mixture model. Then, the generative and discriminant information implied in the FV is exploited after the extraction of the geometrical information between neighboring pixels within different scales. Two experiments are conducted to demonstrate the performance of the proposed MLFP, which are high-resolution SAR image land-cover classification and SAR scene classification. Compared with some existing approaches, the proposed MLFP shows superior performance after the statistical and spatial structure information from different land covers is exploited.