ONNIS-GI: uma rede neural oscilatória para segmentação de imagens implementada em arquitetura maciçamente paralela

The present thesis aims at proposing a new oscillatory neural network and at implementating using a massively parallel architecture, which is suitable to the implementation of digital vision chips for image segmentation. The final version of the proposed network called ONNIS-GI (Oscillatory Neural Network for Image Segmentation with Global Inhibition) was inspired on a network called LEGION (Locally Excitatory Globally Inhibitory Oscillator Network), also of recent conception. Primarily, an introduction to the procedures of image segmentation is presented with the objective to situate and emphasize the importance of the subject in the general context of artificial vision. The thesis also covers the utilization of artificial neural networks in image segmentation, with emphasis on the oscillatory neural networks, which have presented encouraging results in this area. The implementation of vision chips, integrating image sensors and massively parallel processor networks, is also mentioned emphasizing the practical application of the new network proposed. In the study of the LEGION network, results of original applications in image segmentation are presented, in which its segment temporal separation property is verified. The continuous version of this network, a two-dimensional array of neurons based on differential equations, presents both high computational complexity to implementation using digital hardware and many parameters with little practical setting procedures. On the other hand, its massively parallel topology is particularly adequate to the implementation of analog vision chips suitable to segment images. Based on the good results obtained with the developed applications, a new oscillatory neural network is proposed in two versions, called ONNIS and ONNIS-GI, supplanting the LEGION network in several aspects related to practical implementation. The structure of the processing elements of the two versions of the proposed network, its implementation in massively parallel architecture and results concerning simulations and FPGA implementation are presented, demonstrating the viability of the proposal. As a final result, it is concluded that the ONNIS-GI network is more attractive for practical applications, being an innovative approach, efficient in solving image segmentation problems, and possessing the ability to separate the image segments in time for easier identification. Under a practical point of view, the new network can be used for the implementation of vision chips with massively parallel architecture using digital technology, exploring the speed of such implementations and also presenting flexibility to implement more sophisticated segmentation procedures.