FIGURE 7 - available via license: Creative Commons Attribution 4.0 International
Content may be subject to copyright.
Source publication
Deep convolutional neural network (DCNN) based image codecs, consisting of encoder, quantizer and decoder, have achieved promising image compression results. The major challenge in learning these DCNN models lies in the joint optimization of the encoder, quantizer and decoder, as well as the adaptivity to the input images. In this paper, we propose...
Context in source publication
Similar publications
An important role is played by auto encoders in un-supervised learning for transferring of knowledge, making compressed features and other tasks. This study focuses on the use clustering techniques to identify different patterns in the data set. Use of neural network techniques to learn about new compressed features which can show good patterns in...
Citations
... The major intention behind the image compression is to transmit the images with a lower count of bits [13,33,37]. In the image compression, the redundant bit identification, identification of optimal encoding technique and transformation technique are indeed are the key factors [20,26,38]. The "Joint Photographic Experts Group" is indeed the primary image compression technique that has been developed by the JPEG. ...
Large amounts of storage are required to store the recent massive influx of fresh photographs that are uploaded to the internet. Many analysts created expert image compression techniques during the preceding decades to increase compression rates and visual quality. In this research work, a unique image compression technique is established for Vector Quantization (VQ) with the K-means Linde–Buzo–Gary (KLBG) model. As a contribution, the codebooks are optimized with the aid of hybrid optimization algorithm. The projected KLBG model included three major phases: an encoder for image compression, a channel for transitions of the compressed image, and a decoder for image reconstruction. In the encoder section, the image vector creation, optimal codebook generation, and indexing mechanism are carried out. The input image enters the encoder stage, wherein it’s split into immediate and non-overlapping blocks. The proposed GMISM model hybridizes the concepts of the Genetic Algorithm (GA) and Slime Mould Optimization (SMO), respectively. Once, the optimal codebook is generated successfully, the indexing of the every vector with index number from index table takes place. These index numbers are sent through the channel to the receiver. The index table, optimal codebook and reconstructed picture are all included in the decoder portion. The received index table decodes the received indexed numbers. The optimally produced codebook at the receiver is identical to the codebook at the transmitter. The matching code words are allocated to the received index numbers, and the code words are organized so that the reconstructed picture is the same size as the input image. Eventually, a comparative assessment is performed to evaluate the proposed model. Especially, the computation time of the proposed model is 69.11%, 27.64%, 62.07%, 87.67%, 35.73%, 62.35%, and 14.11% better than the extant CSA, BFU-ROA, PSO, ROA, LA, SMO, and GA algorithms, respectively.
... Compared with autoencoders, stacked autoencoders deepen the network structure and enhance the data representation ability, and the sparseness limitation realizes the function of feature extraction [40], hence, it is widely used in classification and data dimensionality reduction [41]. The structure of SAE is shown in Fig. 13. ...
The unmanned combat aerial vehicle (UCAV) is a research hot issue in the world, and the situation assessment is an important part of it. To overcome shortcomings of the existing situation assessment methods, such as low accuracy and strong dependence on prior knowledge, a data-driven situation assessment method is proposed. The clustering and classification are combined, the former is used to mine situational knowledge, and the latter is used to realize rapid assessment. Angle evaluation factor and distance evaluation factor are proposed to transform multi-dimensional air combat information into two-dimensional features. A convolution success-history based adaptive differential evolution with linear population size reduction-means (C-LSHADE-Means) algorithm is proposed. The convolutional pooling layer is used to compress the size of data and preserve the distribution characteristics. The LSHADE algorithm is used to initialize the center of the mean clustering, which overcomes the defect of initialization sensitivity. Comparing experiment with the seven clustering algorithms is done on the UCI data set, through four clustering indexes, and it proves that the method proposed in this paper has better clustering performance. A situation assessment model based on stacked autoencoder and learning vector quantization (SAE-LVQ) network is constructed, and it uses SAE to reconstruct air combat data features, and uses the self-competition layer of the LVQ to achieve efficient classification. Compared with the five kinds of assessments models, the SAE-LVQ model has the highest accuracy. Finally, three kinds of confrontation processes from air combat maneuvering instrumentation (ACMI) are selected, and the model in this paper is used for situation assessment. The assessment results are in line with the actual situation.
... These image compression methods use manual conversion and quantization methods. Typically, these compression methods damage the edges and textures of the image (Lu et al. 2019). ...
... In Eq. (18), the objective and competency functions are used, and any algorithm that minimizes the value of this objective function is a more appropriate method for compressing images. In Fig. 10, the value of the objective function in the proposed method and the other methods used in the Lu et al. (2019) are compared. ...
Compressing the image causes less memory to be used to store the images. Compressing images increases the transmission speed of compressed images in the network. Vector quantization (VQ) is one of the image compression methods. The challenge of the vector quantization method for compression is the non-optimization of the codebooks. Codebook optimization increases the quality of compressed images and reduces the volume of compressed images. Various methods of swarm intelligence and meta-heuristics are used to improve the vector quantization algorithm, but using meta-heuristic methods based on mathematical sciences has less history. This paper uses an improved sine–cosine algorithm (SCA) version to optimize the vector quantization algorithm and reduce the compression error. The reason for using the SCA algorithm in image compression is the balance between the search for exploration and exploitation search by sine and cosine functions, which makes it less likely to get caught in local optima. The proposed method to reduce the calculation error of the SCA algorithm uses spiral trigonometric functions and a new mathematical helix. The proposed method searches for optimal solutions with spiral and snail searches, increasing the chances of finding more optimal solutions. The proposed method aims to find a more optimal codebook by the improved version of SCA in the VQ compression algorithm. The advantage of the proposed method is finding optimal codebooks and increasing the quality of compressed images. The proposed method implementing in MATLAB software, and experiments showed that the proposed method's PSNR index improves the VQ algorithm's ratio by 13.73%. Evaluations show that the proposed method's PSNR index of compressed images is higher and better than PBM, CS-LBG, FA-LBG, BA-LBG, HBMO-LBG, QPSO-LBG, and PSO-LBG. The result shows that the proposed method (or ISCA-LBG) has less time complexity than HHO and WOA compression algorithms.
... Hyperprior models have been the basis for several subsequent advances with further improvements for density modeling, such as joint autoregressive models , Gaussian mixture/attention (Cheng et al., 2020), and channel-wise auto-regressive models (Minnen & Singh, 2020). Another line of work has proposed to use vector quantization with histogram-based probabilities for image compression (Agustsson et al., 2017;Lu et al., 2019). Contrary to VQ-VAE models, these models typically optimize the rate (or a surrogate of the rate) directly and may include a spatial component for the quantized vectors. ...
Recent neural compression methods have been based on the popular hyperprior framework. It relies on Scalar Quantization and offers a very strong compression performance. This contrasts from recent advances in image generation and representation learning, where Vector Quantization is more commonly employed. In this work, we attempt to bring these lines of research closer by revisiting vector quantization for image compression. We build upon the VQ-VAE framework and introduce several modifications. First, we replace the vanilla vector quantizer by a product quantizer. This intermediate solution between vector and scalar quantization allows for a much wider set of rate-distortion points: It implicitly defines high-quality quantizers that would otherwise require intractably large codebooks. Second, inspired by the success of Masked Image Modeling (MIM) in the context of self-supervised learning and generative image models, we propose a novel conditional entropy model which improves entropy coding by modelling the co-dependencies of the quantized latent codes. The resulting PQ-MIM model is surprisingly effective: its compression performance on par with recent hyperprior methods. It also outperforms HiFiC in terms of FID and KID metrics when optimized with perceptual losses (e.g. adversarial). Finally, since PQ-MIM is compatible with image generation frameworks, we show qualitatively that it can operate under a hybrid mode between compression and generation, with no further training or finetuning. As a result, we explore the extreme compression regime where an image is compressed into 200 bytes, i.e., less than a tweet.
... These image compression methods use manual conversion and quantization methods. Typically, these compression methods damage the edges and textures of the image [4]. ...
Image compression is one of the most attractive and practical topics. Without image compression, the image size becomes too large for storage. The transmission of uncompressing images on computer networks slows down and network bandwidth is wasted. Various approaches to image compression have been proposed so far, one of which is vector quantization using mathematical concepts and image processing to compress images. The LBG algorithm is a practical algorithm for compressing images using vector quantization concepts. Most researchers have used meta-heuristic and optimization algorithms with a modeling approach of swarm behavior of living things to improve the quality of the LBG compression algorithm. This study uses a meta-heuristic method based on sine and cosine algorithms (SCA) to improve the quality of the image compression algorithm. In the proposed mathematical modeling method, the SCA algorithm is improved using spiral equations. The improved SCA algorithm is then used to find the optimal codebook in the LBG compression algorithm. Finding a better codebook in the proposed method will increase the quality of the compression images. The proposed method implemented in MATLAB software and experiments showed that the PSNR index in the proposed method improve the ratio of the LBG algorithm by about 13.73%. Evaluations show that the PSNR index of compressed images in the proposed method is higher and better than PBM, CS-LBG, FA-LBG, BA-LBG, HBMO-LBG, QPSO-LBG, PSO-LBG. The result shows the proposed method (or ISCA-LBG) has less time complexity than HHO and WOA compression algorithms.
... In the study [20], they provided a deep Vector Quantization Network for image compression. Deep convolutional neural network (DCNN) codecs, consisting of an encoder, quantizer, and decoder, have achieved promising image compression results. ...
... In the implementation of the proposed method for image compression, standard and gray images of LENA, BABOON, PEPPERS, BARB, and GOLDHILL is using. Some of these images is using to compress and analyze images are shown in Fig. 6 [20]. Sample images are 512 by 512 in size and are also used for evaluation in most studies. ...
... LENA, BABOON, PEPPERS, BARB, and GOLDHILL standard images are used to evaluate the proposed method. The proposed method compared with compression methods such as LBG, FA-LBG (Firefly algorithm-LBG), BA-LBG (BAT algorithm-LBG), DE-LBG (Differential evolution algorithm-LBG), IPSO-LBG (Improved PSO-LBG), and IDE-LBG (Improved DE-LBG) [20]. Comparison of performance time, PSNR, and competency indicators are using. ...
Today, much of the information is storing in images. To transfer information in the form of images, image compression is required. Compressing images reduces the size of images and sends them faster over the network. One of the most methods of image compression is the vector quantization. For vector quantization compression, the codebook is using in cryptography and decryption. The vector quantization compression method typically uses codebooks that are not optimized, which reduces the compression quality of the images. Choosing the optimal codebook makes compression of images with higher quality. Choosing the optimal codebook is a difficult optimization problem and therefore requires intelligent algorithms to solve it. In this paper, the whale optimization algorithm is used to find the optimal codebook in image compression. Whale Optimization Algorithm has different search strategies and is an ideal algorithm for finding the optimal codebook in images compression. Implementation of the proposed algorithm for compression on several standard images shows that the proposed method compresses images with appropriate quality. The proposed method performs more efficient compression than the proposed algorithms such as particle swarm optimization, bat, and firefly algorithms. The signal-to-noise ratio of the proposed method is higher than the compared methods. Experiments on a set of standard images show the proposed method compared to the Fire Fly, Bat, and Differential evolution, Improved Particle Swarm Optimization, and Improved Differential Evolution methods with a compression execution time of 60.48% and 10.21, respectively. %, 4.79%, 5.09% and 3.94% decreased. The proposed method in compression has a higher PSNR index of about 17% than the Linde-Buzo-Gray method.
... Quantization is used in conjunction with Deep Neural Networks (DNNs) for tasks such as classification [49], incremental learning [49], zero-shot learning [28], generation [34], compression [30] and data retrieval [4]. The discrete quantized feature representations can be used post-hoc [11,34,39] or as a learning objective (i.e. ...
Quantization has been applied to multiple domains in Deep Neural Networks (DNNs). We propose Depthwise Quantization (DQ) where $\textit{quantization}$ is applied to a decomposed sub-tensor along the $\textit{feature axis}$ of weak statistical dependence. The feature decomposition leads to an exponential increase in $\textit{representation capacity}$ with a linear increase in memory and parameter cost. In addition, DQ can be directly applied to existing encoder-decoder frameworks without modification of the DNN architecture. We use DQ in the context of Hierarchical Auto-Encoder and train end-to-end on an image feature representation. We provide an analysis on cross-correlation between spatial and channel features and we propose a decomposition of the image feature representation along the channel axis. The improved performance of the depthwise operator is due to the increased representation capacity from implicit feature decoupling. We evaluate DQ on the likelihood estimation task, where it outperforms the previous state-of-the-art on CIFAR-10, ImageNet-32 and ImageNet-64. We progressively train with increasing image size a single hierarchical model that uses 69% less parameters and has a faster convergence than the previous works.
... Then, the threshold system is used to emit coefficients. In Lu et al. [13] a Deep Convolutionary Neural Network (DCNN) was introduced to encode, quantize and decode image compression. A Fully Convolutionary VQ framework (VQNet) is developed in particular to measure the feature vectors in which representative VQNet vectors are optimised with alternate system parameters. ...
... In DPIO mechanism, map and compass operator as well as landmark operators are employed to generate novel solution, however the semantic of exemplar component p j and visiting city list nc in minus operator (13) are based on the pigeon state. When the pigeon is involved in first stage, then p j defines the j-th component of personal best solution of arbitrarily decided pigeon and Full strategy has been applied nc. ...
... Quantization is a widely adopted method based on the parameter sharing approach for compressing neural networks. It groups parameters into multiple clusters, and replaces all parameters in the same cluster with a representative value [25]. As a consequence, the size and computational cost of the model are reduced in return for slightly degraded accuracy. ...
Deploying neural network models to edge devices is becoming increasingly popular because such deployment decreases the response time and ensures better data privacy of services. However, running large models on edge devices poses challenges because of limited computing resources and storage space. Researchers have therefore proposed various model compression methods to reduce the model size. To balance the trade-off between model size and accuracy, conventional model compression methods require manual effort to find the optimal configuration that reduces the model size without significant degradation of accuracy. In this article, we propose a method to automatically find the optimal configurations for quantization. The proposed method suggests multiple compression configurations that produce models with different size and accuracy, from which users can select the configurations that suit their use cases. Additionally, we propose a retraining method that does not require any labeled datasets for retraining. We evaluated the proposed method using various neural network models for classification, regression and semantic similarity tasks, and demonstrated that the proposed method reduced the size of models by at least 30% while maintaining less than 1% loss of accuracy. We compared the proposed method with state-of-the-art automated compression methods, and showed that it can provide better compression configurations than existing methods.
... Vector quantization (VQ) is a powerful method in the field of data compression because of its promising compression performance and simple decoding structure. It is a learning-based approach mainly used in image compression, [69][70][71][72][73] which has also been utilized in neural signal compression. 60,74 However, the spike reconstruction quality of two VQ-based discriminative compression algorithms presented in Ref. 74 has not been evaluated under low signal-to-noise ratio (SNR) situations. ...
Implantable high-density multichannel neural recording microsystems provide simultaneous recording of brain activities. Wireless transmission of the entire recorded data causes high bandwidth usage, which is not tolerable for implantable applications. As a result, a hardware-friendly compression module is required to reduce the amount of data before it is transmitted. This paper presents a novel compression approach that utilizes a spike extractor and a vector quantization (VQ)-based spike compressor. In this approach, extracted spikes are vector quantized using an unsupervised learning process providing a high spike compression ratio (CR) of 10-80. A combination of extracting and compressing neural spikes results in a significant data reduction as well as preserving the spike waveshapes. The compression performance of the proposed approach was evaluated under variant conditions. We also developed new architectures such that the hardware blocks of our approach can be implemented more efficiently. The compression module was implemented in a 180-nm standard CMOS process achieving a SNDR of 14.49[Formula: see text]dB and a classification accuracy (CA) of 99.62% at a CR of 20, while consuming 4[Formula: see text][Formula: see text]W power and 0.16[Formula: see text]mm2 chip area per channel.
