A Novel Parallel Convolutional Network Architecture for Depth-Dependent Object Recognition

This article presents the evaluation of a novel parallel convolutional neural network, oriented to recognize objects at different distances, in order to find a solution to the problem of variability in the value of confidence with which an object is recognized, by varying the distance of capture of the image with respect to the object. In order to test its performance, two additional convolutional neural network architectures are implemented, a conventional one with multiple branches of identification in parallel and a Directed Acyclic Graph convolutional network with the same parameters as the proposed one, which differ in the training database used and the structure of the network's output. This problem is identified when trying to develop assistive robotic systems that should recognize a particular object in a group of objects in order to be taken by an end effector capable of changing trajectories, avoiding possible collisions in human-machine work environments. Here, four different tools must be recognized at four distances (20, 40, 60 and 80 cm), where the conventional CNN obtain the lowest accuracy (80.6%), while, in comparison between the DAG and the parallel CNN, although their performances have been close, the proposed architecture has obtained better results, with 93% average accuracy. It is concluded that this network is able to function in environments with dynamic reference positions of the objects, allowing being implemented in mobile agents that require relocating an object and establishing a new path given an obstacle to evade.