Image Multi-label Classification (IMC) assigns a label or a set of labels to an image. The big demand for image annotation and archiving in the web attracts the researchers to develop many algorithms for this application domain. The Multi-Instance Multi-Label Learning (MIML) is an important type of machine learning framework proposed recently for IMC. In this framework, an image is described with many regions or instances and can be assigned to multiple categories or labels. The MIML is a single label learning transformation. There are 2 ways to do this transformation: The first one transforms Multi-label to single label. This transformation is called Multi Instance Single-Label Learning (MISL) and applies Multi instance learner to have Single Instance Single-Label Learning (SISL). The second one transforms multi instance to single instance. This transformation is called Single-Instance Multi-Label Learning (SIML) and applies Multi-Label learner to have SISL. We are interested in this paper in the second transformation as a Multi-label problem. Two most important existing techniques are proposed for this transformation (MIML-BOOST and MIML-SVM). The drawbacks of these existing methods that they did not take into considerations: a) the description of the elementary characteristics from the image, b) the correlation between labels. In this paper, we propose a novel algorithm (MIML-GABORLPP), which simultaneously handles these limitations. The algorithm uses Gabor filter bank as feature descriptor to handle the first limitation. It applies the Label Priority Power-set as Multi-label transformation to solve the problem of label correlation. The experimental work shows that the results of MIML-GABORLPP are better in terms of four evaluation metrics (Hamming Loss (HL), One Error (OE), Ranking Loss (RL) and Average Precision (AP)) when compared to two other existing techniques.