Study On Objects Recognition Algorithm Based On Gradient Features And Location Information

Objects recognition is an important branch in the field of computer vision, after decadesof researchers efforts, objects recognition technology has been used in the many fields aroundus successfully and widely, it have achieved good results in agriculture, medical services andintelligent traffic control industry, it also plays a very important role in industrial applications,aerospace industry and military fields. An ideal objects recognition system is able to simulatethe human brain to perceive the visual scene and identify the objects and track it accurately,the complex scenes and the learning ability of the cognitive system make this technology bestill in continuous development stage. When we recognize a specific object, the most commonmethod is to use feature extraction and combine of classifier training. Therefore, this paperproposed objects recognition algorithm based on gradient features and location information,that is combine with composite solids location information when extract the feature, so as toimprove the recognition accuracy.In image preprocessing, we present a kind of image segmentation based on the region ofcolor space and LBP algorithm, it is an improvement of MSRM algorithm, it reduces thenumber of regional dimension feature vectors effectively, and more better than the MSRMalgorithm in execution efficiency, and improve the feature extraction accuracy of training set.In the model of object feature extraction, we improve the traditional feature extractionmethods, that is we divide the whole object into several composite entities, and extract theobject gradient feature,here we use HOG algorithm, it use a sliding window mechanism toextract the object appearance edge feature, first normalized color space to reduce the impactof light and the background environment, then calculation gradient and gradient histogramcalculation in the cell, and normalize the block gradient histogram composed by cells, finallyextract the feature vectors, then extract the location information of each composite entity tocreate a training image set of features library. In the model of classifier training, we use SVM and Boosting classification algorithmswhich were used commonly. SVM has better generalization performance and optimizeperformance for different environments of objects recognition and detection; Boostingmethod is simple and robust, fast, can deal with background noise and occlusion effectivelyand it is hot in image understanding, object recognition and other fields.Finally, the experiment showed that objects recognition algorithm based on gradientfeatures and location information what this paper present can eliminate error sampleseffectively and the precision and recall rate is improved, the overall performance is better thanthe original algorithm.