Evaluating the Performance of RCNN Variants for Multi-Object Analysis in Images

Object recognition remains a key challenge in computer vision and robotics. CNNs, originating in neuroscience, have become essential for object identification and recognition. The structural and functional similarities between CNNs and the brain’s visual cortex make them highly effective for developing and testing object classification and identification algorithms. Although standard CNNs typically use a feed-forward structure, the biological visual system utilizes Recurrent Convolutional Neural Networks (RCNNs), which integrate recurrent connections into each convolutional layer. These recurrent connections facilitate bidirectional information flow within the network, allowing each unit in the intermediate layers to dynamically adapt its activity based on the activity of its neighbouring units. This contextual processing is essential for reliable object recognition. This study implements and assesses several leading RCNN-based object detection frameworks - namely, RCNN, Fast RCNN, and Faster RCNN - for the task of identifying mechanical vehicle parts. A customized dataset of vehicle components is employed for training and evaluation, facilitating a direct comparison of the performance of various architectures within this particular application.