Perception of Natural Scenes: Objects Detection and Segmentations using Saliency Map with AlexNet

Abstract

Object detection and classification play a crucial role in accurately tracking objects in complex environments. In recent years, there has been a significant increase in interest among researchers towards object analysis, fueled by the necessity to address challenges and explore opportunities across diverse technological domains. This study introduces a methodologically novel method for image classification through a custom-designed architecture inspired by AlexNet, tailored to process feature vectors for improved pattern recognition. The methodology incorporates Density-Based Spatial Clustering of Applications with Noise (DBSCAN) segmentation to partition images into meaningful regions, showcasing computational efficiency. Additionally, saliency mapping highlights visually significant areas within these segmented images. Various feature extraction methods, including Maximally Stable Extremal Regions (MSER), Binary Robust Invariant Scalable Keypoints (BRISK), and Wavelet transform, are employed to capture unique structures within the images. These features are then fused and optimized using the Fish Swarm Algorithm (FSA), a nature-inspired optimization technique. The refined features, enhanced through the FSA process, are input into a modified AlexNet architecture, enhancing image classification accuracy. The evaluation metrics used include accuracy, precision, recall, and F1-score, providing a comprehensive assessment of performance. The proposed model achieved a classification accuracy of 95.65% on the VOC 2012 dataset, outperforming contemporary methods by a margin of 2-5%, and 93.66% and 92.71% on Caltech-101 and Microsoft Common Objects in Context (MS COCO) datasets, respectively. This innovative blend of techniques harnesses the strengths of FSA and deep learning, yielding precise and robust classification outcomes, outperforming many contemporary methods on datasets like VOC 2012, Caltech 101, and MS COCO.