The scheme of Content-Based Image Retrieval (CBIR) is developed for the process of retrieving of images on the basis of the visual elements defining the image content. Even though, it has certain complications of huge calculation problems such as scoring of the image similarity and feature extraction. The conventional feature extraction models only aim toward the high or low-level features and utilize few manmade features to minimize these gaps. It is required to establish a framework for feature extraction to minimize the gap without utilizing the manmade features from joining low and high-level features. The deep learning is considered as a very powerful tool for feature indication that can represent high and low-level information entirely. In this paper, new optimal deep learning model-based feature extractions for CBIR model are performed for retrieving the accurate images. In the training phase, diverse images are gathered from the standard database, which is used into Optimal Deep Feature Selection Model (ODFSM)by the deep learning models with VGG16, VGG19, Inception, Xception, Resnet50, Resnet101, Resnet 152, and Densenet for extracting the features. Here, the optimization in deep learning models takes place with the help of hybrid optimization algorithm known as Probability-based Coyote-Forest Optimization Algorithm (P-CFOA) for getting the efficient features. In the testing phase, the query images are given into optimal deep feature selection model for extracting the features of the query image. Then, the weighted multisimilarity is performed between the extracted features of database images and query image, in which the weights are optimized using the same P-CFOA. The database images with less weighted multi-similarity when correlated with extracted features of query image are retrieved finally. The simulation analyses are made to demonstrate the better efficiency of image retrieval through the optimal deep feature extractors.