Embedded Descriptor Carriers Computation using Multi-Layer Neural Networks on Large Datasets

Abstract

Intelligent and effective visual extraction due to extensive data sets is unavoidable need included in today. Raw image labels must correspond to visual characteristics in order to extract images based on content based image retrieval (CBIR). CBIR is extensive method progressively applied on retrieval methods. CNN major job is to retrieve authentic and useful images. Various methods have been employed to enhance the effectiveness and reliability of image exploration, such as filename-based searches and image tagging. However, these techniques have not proven successful in real-world applications. To effectively categorize images and function as a filter, the feature vector must include comprehensive visual information. This information should encompass elements such as color, shape, objects, and different types of spatial data. By incorporating these details, the feature vector can more accurately define the image's category and improve the overall efficiency of image exploration. The proposed method excels at detecting, describing, recognizing, and correlating image signatures that accurately reflect the true content of an image. It achieves this by categorizing semantic groupings of nearly identical images. This technique is particularly effective during image retrieval and feature detection processes. The provided methodology details a convolutional neural network (CNN) based method for colorizing grayscale images. The approach begins with defining the area of interest and potentially converting the image to grayscale. Pixel intensities are compared, and patterns within the image are identified using techniques like concentric, retinal, or log-polar methods. Selective pixel sampling, differentiation to analyze neighboring pixels, and smoothing to reduce noise are all employed. Convolution and pooling operations further refine the data. Activation functions, like ReLU or SoftMax, are then applied. Finally, fully connected layers likely within a neural network come into play. The later stages involve sample collection, redundancy measures, distance calculations, and potentially techniques like Bag-of-Words (BoW) and K-Nearest Neighbors (KNN) for image classification. An FV is aggregated, and Bow, KNN, and results are generated. The presented method is capable of identifying, characterizing and correlating images signatures that precisely represent the actual content of a picture by categorizing semantically grouped, nearly identical images. The proposed technique occur during image retrieval and feature detection. Extensive experimentations are conducted on highly recognized datasets such as ALOT-250 and 17-Flowers with mismashed of RestNet-50, Inception and VGG-19. Remarkable results indicated that the presented technique demonstrates significant precision rate and recall rate for huge image groups of challenging datasets.