Le Dong

Hui Li, Pengfei Yang, Juanyang Chen et al.

Knowledge distillation as an efficient knowledge transfer technique, has achieved remarkable success in unimodal scenarios. However, in cross-modal settings, conventional distillation methods encounter significant challenges due to data and statistical heterogeneities, failing to leverage the complementary prior knowledge embedded in cross-modal teacher models. This paper empirically reveals two critical issues in existing approaches: distillation path selection and knowledge drift. To address these limitations, we propose MST-Distill, a novel cross-modal knowledge distillation framework featuring a mixture of specialized teachers. Our approach employs a diverse ensemble of teacher models across both cross-modal and multimodal configurations, integrated with an instance-level routing network that facilitates adaptive and dynamic distillation. This architecture effectively transcends the constraints of traditional methods that rely on monotonous and static teacher models. Additionally, we introduce a plug-in masking module, independently trained to suppress modality-specific discrepancies and reconstruct teacher representations, thereby mitigating knowledge drift and enhancing transfer effectiveness. Extensive experiments across five diverse multimodal datasets, spanning visual, audio, and text, demonstrate that our method significantly outperforms existing state-of-the-art knowledge distillation methods in cross-modal distillation tasks. The source code is available at https://github.com/Gray-OREO/MST-Distill.

Le Dong, Qixuan Cao, Lei Pu et al.

ERVD: An Efficient and Robust ViT-Based Distillation Framework for Remote Sensing Image Retrieval

Le Dong, Ling He, Gaipeng Kong et al.

In this paper, we propose a compact network called CUNet (compact unsupervised network) to counter the image classification challenge. Different from the traditional convolutional neural networks learning filters by the time-consuming stochastic gradient descent, CUNet learns the filter bank from diverse image patches with the simple K-means, which significantly avoids the requirement of scarce labeled training images, reduces the training consumption, and maintains the high discriminative ability. Besides, we propose a new pooling method named weighted pooling considering the different weight values of adjacent neurons, which helps to improve the robustness to small image distortions. In the output layer, CUNet integrates the feature maps gained in the last hidden layer, and straightforwardly computes histograms in non-overlapped blocks. To reduce feature redundancy, we implement the max-pooling operation on adjacent blocks to select the most competitive features. Comprehensive experiments are conducted to demonstrate the state-of-the-art classification performances with CUNet on CIFAR-10, STL-10, MNIST and Caltech101 benchmark datasets.

Gaipeng Kong, Le Dong, Wenpu Dong et al.

This paper addresses the problem of large-scale image retrieval. We propose a two-layer fusion method which takes advantage of global and local cues and ranks database images from coarse to fine (C2F). Departing from the previous methods fusing multiple image descriptors simultaneously, C2F is featured by a layered procedure composed by filtering and refining. In particular, C2F consists of three components. 1) Distractor filtering. With holistic representations, noise images are filtered out from the database, so the number of candidate images to be used for comparison with the query can be greatly reduced. 2) Adaptive weighting. For a certain query, the similarity of candidate images can be estimated by holistic similarity scores in complementary to the local ones. 3) Candidate refining. Accurate retrieval is conducted via local features, combining the pre-computed adaptive weights. Experiments are presented on two benchmarks, \emph{i.e.,} Holidays and Ukbench datasets. We show that our method outperforms recent fusion methods in terms of storage consumption and computation complexity, and that the accuracy is competitive to the state-of-the-arts.

Le Dong, Na Lv, Qianni Zhang et al.

This paper describes an effective and efficient image classification framework nominated distributed deep representation learning model (DDRL). The aim is to strike the balance between the computational intensive deep learning approaches (tuned parameters) which are intended for distributed computing, and the approaches that focused on the designed parameters but often limited by sequential computing and cannot scale up. In the evaluation of our approach, it is shown that DDRL is able to achieve state-of-art classification accuracy efficiently on both medium and large datasets. The result implies that our approach is more efficient than the conventional deep learning approaches, and can be applied to big data that is too complex for parameter designing focused approaches. More specifically, DDRL contains two main components, i.e., feature extraction and selection. A hierarchical distributed deep representation learning algorithm is designed to extract image statistics and a nonlinear mapping algorithm is used to map the inherent statistics into abstract features. Both algorithms are carefully designed to avoid millions of parameters tuning. This leads to a more compact solution for image classification of big data. We note that the proposed approach is designed to be friendly with parallel computing. It is generic and easy to be deployed to different distributed computing resources. In the experiments, the largescale image datasets are classified with a DDRM implementation on Hadoop MapReduce, which shows high scalability and resilience.

Le Dong, Xiuyuan Chen, Mengdie Mao et al.

This paper proposes a classification network to image semantic retrieval (NIST) framework to counter the image retrieval challenge. Our approach leverages the successful classification network GoogleNet based on Convolutional Neural Networks to obtain the semantic feature matrix which contains the serial number of classes and corresponding probabilities. Compared with traditional image retrieval using feature matching to compute the similarity between two images, NIST leverages the semantic information to construct semantic feature matrix and uses the semantic distance algorithm to compute the similarity. Besides, the fusion strategy can significantly reduce storage and time consumption due to less classes participating in the last semantic distance computation. Experiments demonstrate that our NIST framework produces state-of-the-art results in retrieval experiments on MIRFLICKR-25K dataset.