Image retrieval is a fundamental task in computer vision, aims to retrieve relevant images from large-scale databases based on user queries. However, the existing image retrieval systems are their susceptibility to inaccuracies when dealing with semantic gaps among low-level and high-level features leading to mismatches in retrieved results. This work proposes a novel Deep learning (DL) based DEEP-FIR approach to enhance image retrieval efficiency by integrating bio-inspired optimization and deep learning network. Initially, the input query images are pre-processed with weighted median filter to remove the noisy distortions. The proposed method employs an advanced DL-based Regression network that extracts the low-level features like colour and texture, with high-level semantic features. By fusing extracted features, the retrieval system is the leading to more accurate and discriminative representations of images. Additionally, butterfly mating optimization (BMO) algorithm is utilized for boosting performance by calculating comparison between query and database images to specific retrieval tasks. Experimental results on benchmark datasets establish the efficiency of the proposed DEEP-FIR approach with an overall accuracy of 97.8%. The proposed DEEP-FIR approach increases the overall accuracy of 3.16%, 2.35%, and 4.70% for DL-CNN, Multi-view and CBIR-CNN respectively.