Design Of Compound Eye Identification System Based On Deep Learning

With the continuous improvement of artificial intelligence technology,there are more and more demands for robots with visual ability in military and civil fields.The current machine vision platform is dominated by monocular detection system architecture.While monocular detection system only has a small field of view,it can not complete the field of view coverage and comprehensive image capture in a wide range of environments,resulting in low detection flexibility.At present,the target recognition algorithm based on machine vision platform is mainly realized by the method of artificial target feature definition and matching.In the process of defining target feature,the accuracy of target recognition often decreases due to human error,and it is almost impossible to precisely define complex targets.So the monocular detection system and the artificially defined feature matching target recognition algorithm are not suitable for the use of modern machine vision environment.In order to solve these problems,also improve the detection range and target recognition ability of machine vision system,through the research of insect compound eye structure and the construction of neural network,designed a bionic compound eye identification detection system based on deep learning is designed.The system collects image information containing target in a wide range by multi-aperture camera imaging method,integrates the information and uses compressed YOLO V3 deep convolution neural network as the main algorithm for target recognition.At the same time,KCF target tracking algorithm is used to lock the identified target,so as to achieve real-time target recognition and track.Experiments show that the system can recognize and track distant targets in a wide field of view,and solve the problem that large field of view and long observation distance can not be combined in previous target detection systems.Moreover,the pattern recognition algorithm and neural network are combined to ensure the target recognition rate and flexibility,so that the system can learn and recognize a variety of targets.It also reduces the occupancy space of the neural network framework to the computing environment,and provides support for additional path planning,target location and other subsequent algorithms.