Research On Probability Statistical Model For Image Classification

Image classification is a fundamental application for computer vision.With the ad-vent of the technology era,the number of digital image has grown rapidly and posed a huge challenge to automatic image recognition.In order to obtain more accurate im-age recognition results,many image classification algorithms based on statistical model have been proposed in recent years.Among them,the probability statistical model has obtained outstanding performance and been widely used.It constructs the relationship among low-level features to learn abstract features.Since these features fix the semantic gap between the category label and the low-level feature,methods based on probability statistical model have obtained more accurate results for image classification.Probability statistical models have two important branches:graphical model and deep neural net-work.The former learns the latent topics in the image and deep neural network applys an end-to-end structure to automatically learn discriminative features.There are more images with complex background emerging online,meanwhile more fine-grained image labels are required.These categories share great appearance similarity.It is an urgent and challenging task to construct effective probability statistical model to extract more discriminative features.In this thesis,we focus on the image classification method based on the probability statistical model.We first explore the influence of location information and class-sharing property for coarse-grained image recognition.Then,the core problems such as how to eliminate the background interference and extract the pose-robust features are discussed.Furthermore,we analyze the influence of text-level feature for fine-graiend image classification.The details are listed as follows.1.Most traditional methods on graphical model use the histogram of visual word,which ignores spatial information.While location information provides useful infor-mation for image classification.We construct a probability statistical model based on location information for image classification.The model divides the image into multi-ple regions.According to the location,color,texture and other low-level features,we assign a latent topic to each region.Meanwhile,we construct a supervised local-spatial-constraint graphical model to learn class-specific latent topics.The variational method is used to optimize the model parameters and the maximum posteriori probability algorithm is applied to predict the label of each image.2.Since the spatial information fails to classify image when there exists large vari-ation of location for the same class,we construct a probability statistical model to use class-sharing property.This model learns local-class-sharing and local-class-specific la-tent topics.After learn the class-sharing feature,the discriminative power of the class-specific topics is enhanced.This method consists of two steps:firstly we learn the la-tent topics from the histogram of visual words,then support vector machine classifier is trained according to the distribution of the latent topic.In addition,the method uses a deep network to learn the convolutional feature,which are quantized to obtain visual words.Then,we apply the graphical model to learn the latent topic.It is found that the model can learn more discriminative middle-level features,which are complementary to the deep features.The combination of graphical model and deep network can further improve the classification performance.3.To eliminate the noise from complex background,we propose a probability sta-tistical model based on saliency to extract object-level feature.This model can simulta-neously perform object annotation and image classification.Firstly,we generate a set of super-pixels and extract local feature inside each region.Then,our graphical model ob-tains the posterior probability distribution and assigns a topic to each region according to the spatial information,local context and saliency value.To further boost the performance of fine-grained classification,we extract features from the foreground and use these fea-tures to train classifier.Finally,the output scores of graphical model and classifier are combined.The top ranked label is chosen as the prediction of the input image.4.Since there exists large variation of pose for fine-grained categories,we propose a probability statistical model to extract pose-robust feature.In the first phase,we generate a set of polygons composed of randomly selected parts.For each polygon,we use the corresponding deep feature to train classifier.Then,a greedy algorithm with tree structure is applied to choose polygon-based classifiers.In the second phase,the confusing classes are selected according to the results in validation set.For these classes,we train a set of polygon-based classifier and apply greedy algorithm to select discriminative classifiers.Given a test image,the classifier trained in the first step is applied to obtain a coarse result.Then,the second-step classifier is used to distinguish the confusing classes.5.Since text-level feature can provide useful information to distinguish stores,we construct a probability statistical model based on text-level feature for fine-grained store classification.Two features are introduced to distinguish different stores:Text-Exemplar-Similarity and Hypotheses-Weighted-CNN.For the former one,we learn a set of discrim-inative text detectors for each class and use the corresponding output score to represent text-level feature.For the second kind of feature,several object hypotheses are firstly generated.Then,we introduce edge boundary and repeatness to caculate the object s-core of each hypothese.The deep features from all hypotheses are fused according to the corresponding scores.Finally,we design an efficient method to optimize the weights of text-level and image-level feature.