Automated image and video object detection based on hybrid heuristic-based U-net segmentation and faster region-convolutional neural network-enabled learning

Object detection is one of the major areas of computer vision, which adopts machine learning approaches in diverse contributions. Nowadays, the machine learning field has been directed through Deep Neural Networks (DNNs) that takes eminent features of progressions in data availability and computing power. In all the cases, the quality of images and videos are biased and noisy, and thus, the distributions of data are also considered as imbalanced and disturbed. Different techniques are developed for solving the abovementioned challenges, which are mostly considered based on deep learning and computer vision. Though, traditional algorithms constantly offer poor detection for dense and small objects and yet fail the detection of objects through random geometric transformations. One of the categories of deep learning called Convolutional Neural Network (CNN) is famous and well-matched method for image-related tasks, in which the network is trained for discovering the numerous features like colour differences, corners, and edges in the images and videos that are combined into more complex shapes. This proposal intends to develop improved object detection in images and videos with the advancements of deep learning models. The three main phases of the proposed object detection model are (a) pre-processing, (b) segmentation, and (c) detection. Once the pre-processing of the image is performed by median filtering approach, the adaptive U-Net segmentation is performed for the object segmentation using the newly proposed Sun Flower-Deer Hunting Optimization Algorithm (SF-DHOA). The maximization of segmentation accuracy and dice coefficient is considered as the main objective of the proposed segmentation. The hybrid meta-heuristic algorithm termed SF-DHOA is proposed with Sun Flower Optimization (SFO) and Deer Hunting Optimization Algorithm (DHOA), which is used for optimally tuning the U-Net by optimizing the encoder depth and the number of epoch. Further, the detection is performed by the modified Faster Region-Convolutional Neural Network (Faster-RCNN), in which the optimization of number of epoch is performed by hybrid SF-DHOA algorithm with the intention of minimizing the error and training loss function. The performance of the proposed algorithm is evaluated, and the proposed algorithm shows high improvement when compared to existing deep learning-based algorithms.