Research On Multi Scale Representation For Image Based Learning

Multi-scale representation of image is an image analysis method that is similar to the human visual system, which at present has arisen widespread concern. Due to the object features perform in a specific range of scales, in different scales, the visual information from the visual perception is different among persons. When processing the image analysis, compared to the analysis method based on the single image, on which multi-scale can be more meet the needs of application. People have advanced the solution of coping image signals as multi-scale and so on currently. For example,the method of wavelet transform, that the image signal projects at multiple scales in frequency space, in which the provided the high or low frequency information can be obtained under different scales. Nowadays, the order to research the method of image representation is pursuit representation more effectively, namely sparse representation.In the fields of image processing and analysis, it can extract inherent feature of image effectively by sparse coding.In this thesis, we put forward methods to represent multi-scale image sparsely and adaptively, which combining the feathers of scale and sparse. The main work and innovation are as follows:First, multi-scale transformation method based on learning. According to non-adaptive problem about wavelet filter, this paper puts forward using convolution sparse coding model to learning the filters which are adaptive for image content,instead of the filter of wavelet transform(the wavelet function and scale function)performs image multi-scale decomposition, in order to obtain more sparse than wavelet transform has the expression. The transform method based on wavelet transform, blended the sparsity of convolutional sparse coding and made more effective than wavelet transform.Second, the multi-scale image representation method based on learning. For some images, such as medical cell images can be seen as which is composed by a number of cells. We call these basic units as pattern. In this thesis, we put forward image sparse representation method in structure and semantic level, using the patterns with different scales learned by convolutional sparse coding to represent images sparsely.The same image can contain multiple target objects. Different in the scale of the image is presented in the target elements is not the same. This thesis proposed using deconvolutional networks with two layers to analysis the sizes of patterns, then the patterns with different scales can be learned by convolutional sparse coding and used to represent images sparsely.