Review on Image Representation Compression and Retrieval Approaches

Damahe, Lalit B.; Thakur, Nileshsingh V.

doi:10.4018/978-1-5225-6164-4.ch009

Cited by 5 publications

(4 citation statements)

References 79 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Content based image retrieval method had mainly focused on three main issues: visual feature, similarity metric and relevance feedback. Several comprehensive reviews had been recently published [1][2][3][4]8].…”

Section: Related Workmentioning

confidence: 99%

“…The previous efforts mainly focused on developing a robust handcraft feature extraction to enhance the performance of the retrieval system [1]. Traditional Remote Sensing Image Retrieval (RSIR) method uses handcraft features to represent the content of the images [1][2][3][4]. These features could be categorized based on features type into global and local.…”

Section: Introductionmentioning

confidence: 99%

“…The key to image retrieval is to design or learn representative descriptions especially for satellite images. Different natural image descriptions have been applied for satellite image retrieval [4]. These conventional descriptions include shape, color, texture and other features.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Learning to Hash with Convolutional Network for Multi-label Remote Sensing Image Retrieval

Moustafa¹,

Ahmed²,

Hamed³

2020

IJIES

View full text Add to dashboard Cite

Recently, deep hashing dominated single label image retrieval approaches. However, the complex nature of remote sensing images, which likely contains multi-labels, hardly benefits from the above approaches. To overcome single-label image retrieval limitations in remote sensing domain, we address this problem by proposing a multi-label remote sensing image retrieval (MLRSIR-NET) framework. Specifically, the proposed MLRSIR-NET composed of two main sub-networks: multi-level feature extraction and deep hash. The multi-level feature extraction network predicts multi-level features to exploit different levels of Convolution Neural Network (CNN (characteristics. To suppress discriminative feature representation, the multi-level features are aggregated and feed to Convolutional Block Attention Module (CBAM) to amplify the representation of relevant multi-label features. CBAM is flexibly integrated into our network with end-to-end training. The hash network stacked two fully connected layers aimed to learn multiple hashing functions to encode the feature vector into a compact hash code. Finally, we conduct experiments on two benchmarks for multi-label images: Dense Labelling Remote Sensing Dataset (DLRSD) and Wuhan Dense Labeling Dataset (WHDLD) to systematically assess the performance. The results show that the proposed framework improved the accuracy in terms of Mean Average Precision (MAP) by a considerable margin of 85.4%, 87.2%, 90.8% and 92.9% for 12-bit, 24-bit, 32-bit and 48-bit code lengths respectively on DLRSD. For WHDLD, it can be noted that the proposed framework supers the DCH by 93.8%, 98.7%, 91.9%, and 94.6% on average respectively.

show abstract

Section: Related Workmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Learning to Hash with Convolutional Network for Multi-label Remote Sensing Image Retrieval

Moustafa¹,

Ahmed²,

Hamed³

2020

IJIES

View full text Add to dashboard Cite

show abstract

“…Albeit succeeding at their tasks, such research endeavors do not emphasize the actual properties of the obtained representations, but rather the model itself. In recent years, the learned representations have been actively studied, offering insights into how existing representations can be compressed for many practical applications, including image [5] and online text classification [6].…”

Section: Introductionmentioning

confidence: 99%

Compressibility of Distributed Document Representations

Škrlj¹,

Petković²

2021

Preprint

View full text Add to dashboard Cite

Contemporary natural language processing (NLP) revolves around learning from latent document representations, generated either implicitly by neural language models or explicitly by methods such as doc2vec or similar. One of the key properties of the obtained representations is their dimension. Whilst the commonly adopted dimensions of 256 and 768 offer sufficient performance on many tasks, it is many times unclear whether the default dimension is the most suitable choice for the subsequent downstream learning tasks. Furthermore, representation dimensions are seldom subject to hyperparameter tunning due to computational constraints. The purpose of this paper is to demonstrate that a surprisingly simple and efficient recursive compression procedure can be sufficient to both significantly compress the initial representation, but also potentially improve its performance when considering the task of text classification. Having smaller and less noisy representations is the desired property during deployment, as orders of magnitude smaller models can significantly reduce the computational overload and with it the deployment costs. We propose CORE, a straightforward, compression-agnostic framework suitable for representation compression. The CORE's performance is showcased and studied on a collection of 17 real-life corpora from biomedical, news, social media, and literary domains. We explored CORE's behavior when considering contextual and non-contextual document representations, different compression levels, and 9 different compression algorithms. Current results based on more than 100,000 compression experiments indicate that recursive Singular Value Decomposition offers a very good trade-off between the compression efficiency and performance, making CORE useful in many existing, representation-dependent NLP pipelines.

show abstract