Content-based image retrieval (CBIR) is a process of finding similar images from a large dataset of images for a given input image. Retrieving comparatively similar images which feature to be considered is an important task in the whole CBIR process. Since the evolution of CBIR, there is a lot of work on the features extraction part so as to reduce the semantic gap and improve image retrieval results. In the initial stage of image retrieval, low-level features like colour, texture, shape, borders, etc. are considered and after the evolution of deep learning, high-level features were extracted from images proved to be better than earlier methods of image retrieval. In this paper, we propose to use features pre-trained CNN model combinations, which are trained for large image database classification. This approach produces superior results of classification and retrieval for various datasets by considering original query images and acceptable results by considering rotated query images.