摘要:
Learning is sequentially executed with respect to weak discriminators based on learning data held in a storage device. Upon learning, an evaluation value for the weak discriminator is calculated. It is discriminated, based on a shift of the evaluation value, whether or not the learning is overlearning. If it is discriminated that the learning is overlearning, new learning data is added. Thus, the overlearning is easily detected and the learning is efficiently executed.
摘要:
An object detecting apparatus includes a plurality of feature value calculating units that are provided for respective different features of an image and perform a process of extracting the features from an attention region in parallel; a plurality of combining units detecting combinations of the features in parallel, the plurality of combining units are provided for the respective combinations of the features included in the attention region, the plurality of combining units detect the combinations from a outputted features from the plurality of feature value calculating units; and a plurality of identifying units that are provided corresponding to the plurality of combining units and perform in parallel a process of identifying an object based on the combinations detected by the combining units.
摘要:
Disclosed are methods and systems, among which is a system that includes a pattern-recognition processor, a central processing unit (CPU) coupled to the pattern-recognition processor via a pattern-recognition bus, and memory coupled to the CPU via a memory bus. In some embodiments, the pattern-recognition bus and the memory bus form about the same number of connections to the pattern-recognition processor and the memory, respectively.
摘要:
A pipeline architecture for analyzing multiple streams of video is embodied, in part, in a layer of application program interfaces (APIs) to each stage of processing. Buffer queuing is used between some stages, which helps moderate the load on the CPU(s). Through the layer of APIs, innumerable video analysis applications can access and analyze video data flowing through the pipeline, and can annotate portions of the video data (e.g., frames and groups of frames), based on the analyses performed, with information that describes the frame or group. These annotated frames and groups flow through the pipeline to subsequent stages of processing, at which increasingly complex analyses can be performed. At each stage, portions of the video data that are of little or no interest are removed from the video data. Ultimately, “events” are constructed and stored in a database, from which cross-event and historical analyses may be performed and associations with, and among, events may be made.
摘要:
Methods, systems and computer program products for simplifying complex data stream problems involving feature extraction from noisy data. Exemplary embodiments include a method for processing a data stream, including applying multiple operators to the data stream, wherein an operation by each of the multiple operators includes retrieving the next chunk for each of set of input parameters, performing digital processing operations on a respective next chunk, producing sets of output parameters and adding data to one or more internal data stores, each internal data store acting as a data stream source.
摘要:
The present approach increases bandwidth by performing at least two functions at the pre-processing level. Specifically, under the present approach, program code is structured so that the segmentation and binarization functions/modules (and optionally a blob analysis function/module) are merged into a single module to reduce memory bandwidth. In addition, each image frame is segmented into a plurality of partitions (e.g., vertical strips) to enhance the reusability of the image data in LS already fetched from main memory. Each partition is then processed by a separate one of a plurality of processing engines, thereby increasing the utilization of all processing engines and allowing the processing engines to maintain good bandwidth.
摘要:
In a boundary active only scheme proposed by the present invention, only a cell in a boundary of region growth is brought into an active mode, and the other cells are brought into a standby mode. The respective cells perform state transition in parallel, and decision of the state transition performed for each clock cycle is not performed in a case where any of the three conditions that none of the adjacent cells is ignited, the cell itself is already ignited, and the cell already belongs to a certain divided region is satisfied. Therefore the number of simultaneously operating cells and that of coupling weight registers are minimized, and control is automatically executed to reduce power consumption.
摘要:
A parallel pattern detection engine (PPDE) comprise multiple processing units (PUs) customized to do various modes of pattern recognition. The PUs are loaded with different patterns and the input data to be matched is provided to the PUs in parallel. Each pattern has an Opcode that defines what action to take when a particular data in the input data stream either matches or does not match the corresponding data being compared during a clock cycle. Each of the PUs communicate selected information so that PUs may be cascaded to enable longer patterns to be matched or to allow more patterns to be processed in parallel for a particular input data stream.
摘要:
In some embodiments, multi-core stochastic discrimination is generally presented. In this regard, a method is introduced comprising providing random regions of a feature space to parallel cores, testing each random region for enrichment in parallel, recording coverage for each data point in each enriched random region in parallel, and calculating an overall average coverage for each data point among the enriched random regions. Other embodiments are also disclosed and claimed.
摘要:
A general purpose processor architecture (methods and apparatuses) that can discern all subsets of a serial data stream which fulfill an arbitrarily complex reference pattern. The invention comprises an ordered set of Detection Cells conditionally interconnected according to the reference pattern and operationally controlling one another's states through the network. The invention preferably includes a Host Interface to enable reporting of Results from a search session as well as the input and control of reference patterns and source data.