67results about How to "Small memory footprint" patented technology

A method and system of lossless compression of integer data using a novel backward-adaptive technique. The adaptive Run-Length and Golomb/Rice (RLGR) encoder and decoder (codec) and method switches between a Golomb/Rice (G/R) encoder mode only and using the G/R encoder combined with a Run-Length encoder. The backward-adaptive technique includes novel adaptation rules that adjust the encoder parameters after each encoded symbol. An encoder mode parameter and a G/R parameter are adapted. The encoding mode parameter controls whether the adaptive RLGR encoder and method uses Run-Length encoding and, if so, it is used. The G/R parameter is used in both modes to encode every input value (in the G/R only mode) or to encode the number or value after an incomplete run of zeros (in the RLGR mode). The adaptive RLGR codec and method also includes a decoder that can be precisely implemented based on the inverse of the encoder rules.

A method and system of lossless compression of integer data using a novel backward-adaptive technique. The adaptive Run-Length and Golomb/Rice (RLGR) encoder and decoder (codec) and method switches between a Golomb/Rice (G/R) encoder mode only and using the G/R encoder combined with a Run-Length encoder. The backward-adaptive technique includes novel adaptation rules that adjust the encoder parameters after each encoded symbol. An encoder mode parameter and a G/R parameter are adapted. The encoding mode parameter controls whether the adaptive RLGR encoder and method uses Run-Length encoding and, if so, it is used. The G/R parameter is used in both modes to encode every input value (in the G/R only mode) or to encode the number or value after an incomplete run of zeros (in the RLGR mode). The adaptive RLGR codec and method also includes a decoder that can be precisely implemented based on the inverse of the encoder rules.

This invention discloses an implementation of audio codec, which has low computational complexity, small memory footprint and high coding efficiency. It can be used in handheld devices, SoC or ASIC products and embedded systems. At the encoder side: first, apply time-to-frequency transform to audio signals, obtaining un-quantized spectrum data; second, based on the un-quantized spectrum data and target bit count, calculate the corresponding information of optimal scale factor, frequency band group, code table index and quantized spectrum by iteration; third, calculate and format bit-stream; fourth, output formatted bit-stream. At the decoder side: parse the formatted bit-stream, apply decoding and inverse quantization to the spectrum of each frame, reconstruct temporal audio data by frequency-to-time transform, and reconstruct the time-domain signals of each channel.

A system and method facilitating pattern recognition is provided. The invention includes a pattern recognition system having a convolutional neural network employing feature extraction layer(s) and classifier layer(s). The feature extraction layer(s) comprises convolutional layers and the classifier layer(s) comprises fully connected layers. The pattern recognition system can be trained utilizing a calculated cross entropy error. The calculated cross entropy error is utilized to update trainable parameters of the pattern recognition system.

The 3D window system utilizes hardware accelerated window system rendering to eliminate the pixel copy step of 3D window system output redirection. The 3D window system includes a window server that directs the window system device driver graphics routines to render into the texture memory of a graphics device.

A hierarchical power verification system and method creates abstract models of power behavior of modules that it successfully verifies. The abstract models simplify the module definition by omitting internal module details but provide sufficient information for power verification of higher level modules that incorporate this abstracted module. Design blocks are replaced with these abstract power models, resulting in reduced run-time and memory requirements. The power models can include power switches inside the block, related supplies of logic ports, supply power states, system power states, power management devices such as isolation logic and level shifters, feed-through and floating ports. The power model may be expressed either in UPF or as a combination of liberty model and UPF. After replacing modules with abstracted models the HPVS can quickly verify an entire SoC with a small memory footprint. When a user modifies a module, the HPVS need only verify the changed module and related modules at higher levels of module hierarchy.

Software atomization provides a method of loading atomized computer program code and data on demand by loading, from an atom database, into a memory, an atom defining code or data in a fine-grained, individually addressable manner. Atoms comprise an atom identifier, computer program code or data information and computer program code or data reference information. The computer program code or data information is modified by transforming the reference information into memory addresses. The present invention provides for the lazy loading of data both implicitly and explicitly. Data is lazily loaded implicitly when it is referenced from code that is lazily loaded. Data is lazily loaded explicitly by encoding a referencing atom identifier such that the referenced data atom is not loaded until actually accessed at runtime.

A method, system, and computer program product of building a Native XML Object Database, the present invention dynamically generates database API from object-oriented design and persists data in native XML files. The structure of data is maintained from front end API to back end file storage for better security and performance.

The present invention is a real-time operating system for use on embedded controllers having minimal resources controlling reactive systems or applications. One application of the present invention is in wireless sensors in a wireless sensor network. By requiring all tasks to have clearly defined start and end points, the present invention eliminates the Waiting state for a task. This allows the RTOS to have a very small memory footprint and for the RTOS to use a single call/return stack.

The invention discloses a stochastic distributed data stream frequent item set mining system. In the system, a new distributed data stream mining mode for frequent item transactions is adopted, a stochastic frequent item mixer is introduced to improve the accuracy of mining, new transactions, when accumulated to the same number as basic windows, in a data steam are send to n frequent item set mining devices by a data item dividing method according to the different items in the transactions respectively, transaction numbers to which different data items belong are attached to the different data items, and in the frequent item set mining devices, frequent item sets are mined by running and operating different frequent item transactions and the mined frequent item sets are stored in a frequent item set storage. Finally, the data items are fed back to data item dividers through the stochastic frequent item mixer for deep mining. Compared with other methods, the method has the advantages of small memory storage space, high response speed and the like. Meanwhile, the coverage of mode mining can be ensured by increasing the number of the frequent item set mining devices or feedback times.

The invention relates to a method for extracting characteristic parameters of a low-voltage area network. The method comprises the following steps: S1, establishing an area network characteristic index system; S2, adopting an association rule mining algorithm-Apriori to realize selection of characteristic parameters of the low-voltage power distribution area; S3, clustering the sample data by adopting an improved K-Means algorithm according to the selected characteristic parameters of the low-voltage power distribution area; and S4, extracting line parameters of the power distribution line byadopting a compact 2D FDFD algorithm. By extracting circuit parameters in the power distribution line, the line loss rate of the system can be reduced by optimizing load power supply, the economic loss is reduced, and the energy conservation and emission reduction work is deeply promoted.

Software atomization provides a method of loading atomized computer program code and data on demand by loading, from an atom database, into a memory, an atom defining code or data in a fine-grained, individually addressable manner. Atoms comprise an atom identifier, computer program code or data information and computer program code or data reference information. The computer program code or data information is modified by transforming the reference information into memory addresses. The present invention provides for the lazy loading of data both implicitly and explicitly. Data is lazily loaded implicitly when it is referenced from code that is lazily loaded. Data is lazily loaded explicitly by encoding a referencing atom identifier such that the referenced data atom is not loaded until actually accessed at runtime.

A method for generating a plurality of indexed data fields based on a pattern set comprising a plurality of patterns, the method comprising detecting for each pattern from the pattern set a pattern offset; creating for each pattern offset in the pattern set an indexed pattern group, wherein the index of the indexed pattern group corresponds to the pattern offset; adding each pattern in the pattern set having the same pattern offset to the indexed pattern group having an index corresponding to the pattern offset; adding each pattern having no specific pattern offset to each of the indexed pattern groups; and compiling each indexed pattern group into an indexed data field.

The invention relates to a WINCE system mirror image constructing method and system, and a WINCE system mirror image. The WINCE system mirror image constructing method comprises the following steps: constructing the nk.nb0 portion of the WINCE system mirror image according to components required for system operation, and storing the nk.nb0 in a hidden space developed by the Bootloader on a system storage device; and constructing the windows.sbi portion of the WINCE system mirror image containing drive and service components except components required in system operation, and releasing the windows.sbi portion to the specified directory of a readable-writable storage device. According to the WINCE system mirror image constructing method and system and the WINCE system mirror image, the nk.nb0 portion and the windows.sbi portion of the WINCE system mirror image are respectively constructed, and the constructed windows.sbi portion is released to the specified directory of the readable-writable storage device, and when the system is started, the nk.nb0 portion is firstly uploaded by the Bootloader, so that the starting speed is enhanced, a small memory space is occupied, the timeliness of the system is ensured, and redundancy backup can be easily implemented.

The invention discloses a line speed flexible multi-tuple message classifying method. The method includes the steps of firstly, converting a port field in a rule into a corresponding prefix closure to obtain a discontinuous mask prefix set, wherein the port field is converted into the corresponding prefix closure, that is, the shortest section used for covering the section of the port field and capable of being represented in the form of prefix is formed, and all dimensions of a five-dimension prefix set are connected together to obtain the discontinuous mask prefix set; secondly, establishing a CMT data structure through the discontinuous mask prefix set; thirdly, conducting high-speed message classification through a CMT structure search algorithm. The line speed flexible multi-tuple message classifying method has the advantages of being high in search speed, small in hardware expense, low in energy consumption and the like.

The invention discloses a trademark pattern matching method and a corresponding device, equipment and medium, and the method comprises the steps of obtaining the image features of a trademark pattern contained in a to-be-compared picture, constructing a to-be-compared feature matrix, and storing the image features of a single trademark pattern in each row vector; obtaining a preset compared feature matrix of a compared trademark library, wherein each row vector stores an image feature of a single compared trademark in the compared trademark library; calculating the similarity based on the to-be-compared feature matrix and the compared feature matrix to obtain a similar matrix; and screening out elements of which the similarity meets a preset condition from the similar matrix so as to determine suspected trademark patterns having similarity association with the compared trademarks from the trademark patterns, and matching a plurality of first compared trademarks most similar to each suspected trademark pattern for each suspected trademark pattern. According to the invention, operation and judgment of approximate matching of the trademark are realized based on matrix operation, the operation efficiency is high, the response is quick, the overhead is relatively low, and the invention is suitable for identifying suspected trademark infringement information from the product picture.