Study On Scene Classification And Recognition Algorithms Of Abnormal Regions In Road Scene

As an important branch of pattern recognition, scene understanding has aroused great attention of researchers from home and abroad. At present, academic researchers most focus on normal images, but the research of anomaly detection in the scene is rarely found. However, in real life, it is often the abnormal phenomena in the scene that may exert detrimental influence on human activities, and even pose a potential threat to public safety. Therefore, scene classification and abnormal scene recognition are both of research value.Scene classification and object recognition are respectively proposed and analyzed in this thesis. For the given scene, anomalies in the picture can be detected and segmented based on the improved algorithms. The main works are as follows:The scene classification algorithm based on deep learning was improved. The representative characteristics can be extracted by deep learning, and new training sets, used for multiple SVM classifiers, can be constructed through cross-validation. Finally, according to a designed vote mechanism, the label of scene images can be predicted through multilevel SVM classifiers. Experiments show that the proposed method is highly precise in classification.A recognition algorithm based on dense matching was given for abnormal images. In a given scene, objects can appear in a limited range. It will be recognized as anomaly scene while objects appear out of that limited range. In order to detect the objects more effectively, a new algorithm based on SIFT-Flow method was adopted in this thesis to describe the local features of the image. Finally, a SVM model can be trained by labeled images through supervised learning, and the anomaly detection of a given scene can be achieved. It is illustrated by experiments that the abnormal object can be detected effectively through adopting the local features and the proposed method is rather accurate in recognition.Selective Search algorithm was adopted to segment the abnormal areas in an image. Through segmenting regions of interest in the scene, classifier can be trained by the features extracted from image patches. Finally, the identification of abnormal areas can be realized. Experiments demonstrate that the algorithm can effectively segment the anomaly in the scene.