An Improved Active Basis Model For Animal Head Image Detection

With the development of computer vision,object detection has become the research basis of many visual tasks.In the field of object detection,the detection of animal head image is always a kind of complicated task because of the complexity of animal head image and the multiple deformable characteristics of the same animal.Active Basis Model is a model with good recognition ability under the condition of learning with few samples,but the detection of animal head image needs to be improved.In the process of Gabor feature extraction,the parameters of Gabor filter cannot be set to ensure the optimal state for each category.If the detection content is changed,the parameter setting of Gabor filter can only find the best scheme through many experiments.In addition,due to the there are some shortcomings of in the matching tracking algorithm used in the learning module of the original model,because of such as its non-orthogonality,which may it brings about repeated iteration and redundant calculation.Sometimes it brings repeated iterations and redundant calculations because of its nonorthogonality.In view of the above situation,this thesis presents an improved model for animal head image detection based on the Active Basis Model,which mainly makes the following two improvements: Firstly,a parameter optimization scheme of Gabor filter is proposed to improve the detection and recognition module of the model.The scheme combines two methods of calculation optimization and experimental verification.On the basis of setting the direction of the filter bank manually,taking the bandwidth parameter as the main parameter,using the constraint relationship between the parameters and the Fisher principle to design the discriminant function,the filter parameters are optimized reasonably by calculation,so that the effect of edge discrimination is better.Secondly,the orthogonal matching tracking algorithm is used to replace the original matching tracking method in the Active Basis Model.Using the characteristics of the orthogonal matching tracking algorithm,that is,all the selected dictionary elements and the dictionary elements to be selected in the residual have orthogonality,so as to prevent the repeated iteration of the selected dictionary.Then the improved tracking method and projection tracking coupling are simplified to a new learning module method.After the improvement of the above two parts,the reasoning module of the original model is used as the verification module of the actual optimization effect.The three modules constitute the improved Active Basis Model.The experimental analysis was carried out by downloading the animal head image data set from the network.The deformable animal head image recognition model in this thesis can complete the task of target head image recognition and position marking when there are only 12 training samples.Compared with the original Active Basis Model,the optimal value of Gabor parameters can be found without previous tests.Through the experimental verification of the template detection effect before and after the improvement,the template F1 score in this thesis is higher,that is,the accuracy of detection is improved.The improvement of the shared sketch learning algorithm makes the learning time shorter.The average time of calculating each small picture in the same device can be reduced by about 0.08 seconds,and the redundant calculation is less.Using fewer Gabor wavelet elements can still show better results.At the end of the thesis,the improved model is used to test the data sets of animal head images.The results show that the model can meet the needs of detection and representation in visual tasks.