Indoor Object Detection using Cosine Adaptive Lightweight Single shot MaGI Detector to assist Blind People

By employing technology to discover and recognize objects tucked away from sight (indoor objects), this technology uses specialized software to gather information from within any type of building or structure (the surrounding environment). These systems collect and analyze data using specialized sensors and software algorithms, to provide the user with immediate feedback. In addition, the system developed in this research paper represents an entirely new approach to what is currently used to help the VI community. It will allow Visually Impaired (VI) people to more easily locate objects around them with a combination of hardware and software known as the Cosine Adaptive Lightweight MaGI Detector (CALMD). The CALMD is a combination of two algorithms, both of which remove noise from natural images and process the data with an enhanced algorithm called the Cosine Adaptive Enhanced Block-Matching and 3D Filtering Network (CA-EBM3D Net). Both use high-frequency filtering techniques and a an averaging method that balance their effects, effectively eliminating noise throughout the image. Using the Cosine Adaptive Lightweight MaGI Detector and RexNet-150 for feature extraction will improve the accuracy and efficiency of indoor bject recognition. Adaptive filtering and cosine similarity are used within the CA-EBM3D Net methodology to achieve improved image denoising whereas the Enhancing Lightweight MaGI Single Shot Detector (ELMSSD) employs a Max-min-Greedy-Interaction (MaGI) optimization framework to identify and prioritize relevant objects within the environment and assist blind users in realtime. This work confirms the effectiveness of the proposed solution in enhancing accessibility and increasing independence for individuals who are visually impaired and illustrates the impact that new technologies can have in building inclusive technology through advanced machine learning methodologies. The proposed method performed exceptionally well in detecting indoor objects, achieving 99.98% accuracy on the dataset, with superior results compared to previous methods. When evaluated using the 10-fold validation technique, the highest accuracy achieved by this method was 99.23%, 99.94% F1-Score, 99.56% Recall, and 99.12% Precision respectively. The state-of-the-art evaluation of this proposed framework reported 98.5% Mean Average Precision (mAP), and 45.56dB Peak Signal-to-Noise Ratio (PSNR) respectively.