In this work, we present a biologically inspired model for recognition of occluded patterns. The general architecture of the model is based on the two visual information processing pathways of the human visual system, i.e. the ventral and the dorsal pathways. The proposed hierarchically structured model consists of three parallel processing channels. The main channel learns invariant representations of the input patterns and is responsible for pattern recognition task. But, it is limited to process one pattern at a time. The direct channel represents the biologically based direct connection from the lower to the higher processing level in the human visual cortex. It computes rapid top-down pattern-specific cues to modulate processing in the other two channels. The spatial channel mimics the dorsal pathway of the visual cortex. It generates a combined saliency map of the input patterns and, later, segments the part of the map representing the occluded pattern. This segmentation process is based on our hypothesis that the dorsal pathway, in addition to encoding spatial properties, encodes the shape representations of the patterns as well. The lateral interaction between the main and the spatial channels at appropriate processing levels and top-down, pattern-specific modulation of the these two channels by the direct channel strengthen the locations and features representing the occluded pattern. Consequently, occluded patterns become focus of attention in the ventral channel and also the pattern selected for further processing along this channel for final recognition.

In this paper a biologically-inspired approach for object search is introduced. This approach is based on the visual information processing in the human brain and more specifically along the two visual processing pathways of the visual cortex. According to this approach different processes, with similar representational structure, work in parallel toward their local tasks, while at the same time, their mutual interaction leads to achievement of larger global goals. The model based on this approach provides a platform where bottom-up and top-down cues are computed and integrated in small incremental steps and lead to emergence of attention that selects an appropriate object. The two important principles of visual information processing, i.e., constraint satisfaction and inhibition play the key role in this model. The model is implemented with an interactive neural network. Simulation results demonstrate the practicality as well as the strength of this approach for object search tasks.

The main focus of this thesis is to develop biologically-based computational models for object recognition. A series of models for attention and object recognition were developed in the order of increasing functionality and complexity. These models are based on information processing in the primate brain, and specially inspired from the theory of visual information processing along the two parallel processing pathways of the primate visual cortex. To capture the true essence of incremental, constraint satisfaction style processing in the visual system, interactive neural networks were used for implementing our models. Results from eye-tracking studies on the relevant visual tasks, as well as our hypothesis regarding the information processing in the primate visual system, were implemented in the models and tested with simulations.

As a first step, a model based on the ventral pathway was developed to recognize single objects. Through systematic testing, structural and algorithmic parameters of these models were fine tuned for performing their task optimally. In the second step, the model was extended by considering the dorsal pathway, which enables simulation of visual attention as an emergent phenomenon. The extended model was then investigated for visual search tasks. In the last step, we focussed on occluded and overlapped object recognition. A couple of eye-tracking studies were conducted in this regard and on the basis of the results we made some hypotheses regarding information processing in the primate visual system. The models were further advanced on the lines of the presented hypothesis, and simulated on the tasks of occluded and overlapped object recognition.

On the basis of the results and analysis of our simulations we have further found that the generalization performance of interactive hierarchical networks improves with the addition of a small amount of Hebbian learning to an otherwise pure error-driven learning. We also concluded that the size of the receptive fields in our networks is an important parameter for the generalization task and depends on the object of interest in the image. Our results show that networks using hard coded feature extraction perform better than the networks that use Hebbian learning for developing feature detectors. We have successfully demonstrated the emergence of visual attention within an interactive network and also the role of context in the search task. Simulation results with occluded and overlapped objects support our extended interactive processing approach, which is a combination of the interactive and top-down approach, to the segmentation-recognition issue. Furthermore, the simulation behavior of our models is in line with known human behavior for similar tasks.

In general, the work in this thesis will improve the understanding and performance of biologically-based interactive networks for object recognition and provide a biologically-plausible solution to recognition of occluded and overlapped objects. Moreover, our models provide some suggestions for the underlying neural mechanism and strategies behind biological object recognition.

In this paper a biologically-inspired model for partly occluded patterns is proposed. The model is based on the hypothesis that in human visual system occluding patterns play a key role in recognition as well as in reconstructing internal representation for a pattern’s occluding parts. The proposed model is realized with a bidirectional hierarchical neural network. In this network top-down cues, generated by direct connections from the lower to higher levels of hierarchy, interact with the bottom-up information, generated from the un-occluded parts, to recognize occluded patterns. Moreover, positional cues of the occluded as well as occluding patterns, that are computed separately but in the same network, modulate the top-down and bottom-up processing to reconstruct the occluded patterns. Simulation results support the presented hypothesis as well as effectiveness of the model in providing a solution to recognition of occluded patterns. The behavior of the model is in accordance to the known human behavior on the occluded patterns.

In this paper a biologically-inspired model for recognition of overlapped patterns is proposed. Information processing in the two visual information processing pathways, i.e., the dorsal and the ventral pathway, is modeled as a solution to the problem. We hypothesize that dorsal pathway, in addition to encoding the spatial information, learns the shape representation of the patterns and, later uses this knowledge as a top-down guidance signal to segment the bottom-up, image-based saliency map. This process of segmentation in the dorsal pathway is implemented as an interactive process, where interaction between bottom-up image information and top-down shape cues lead to incremental development of a segmented saliency map for one of the overlapped object at a time. This segmented map encodes spatial as well as shape information of the respective pattern in the input. The interaction of the dorsal channel with the ventral channel leads to modulation and selective processing of the respective pattern in the ventral pathway for final recognition. Simulation results support the presented hypothesis as well as effectiveness of the model in providing a solution to the recognition of overlapped patterns. The behavior of the model is in accordance to the known human behavior on the occluded patterns.

We present a computational model for visual processing where attentional focus emerges fundamental mechanisms inherent to human vision. Through detailed analysis of activation development in the network we demonstrate how normal interaction between top-down and bottom-up processing and intrinsic mutual competition within processing units can give rise to attentional focus. The model includes both spatial and object-based attention, which are computed simultaneously, and can mutually reinforce each other. We show how a non-salient location and a corresponding non-salient feature set that are at first weakly activated by visual input can be reinforced by top-down feedback signals (centrally controlled attention), and instigate a change in attentional focus to the weak object. One application of this model is highlight a task-relevant object in a cluttered visual environment, even when this object is nonsalient (non-conspicuous).

This thesis deals with biologically-inspired interactive neural networks for the task of object recognition. Such networks offer an interesting alternative approach to traditional image processing techniques. Although the networks are very powerful classification tools, they are difficult to handle due to their bidirectional interactivity. It is one of the main reasons why these networks do not perform the task of generalization to novel objects well. Generalization is a very important property for any object recognition system, as it is impractical for a system to learn all instances of an object class before classifying. In this thesis, we have investigated the working of an interactive neural network by fine tuning different structural and algorithmic parameters.  The performance of the networks was evaluated by analyzing the generalization ability of the trained network to novel objects. Furthermore, the interactivity of the network was utilized to simulate focus of attention during object classification. Selective attention is an important visual mechanism for object recognition and provides an efficient way of using the limited computational resources of the human visual system. Unlike most previous work in the field of image processing, in this thesis attention is considered as an integral part of object processing. Attention focus, in this work, is computed within the same network and in parallel with object recognition.

As a first step, a study into the efficacy of Hebbian learning as a feature extraction method was conducted. In a second study, the receptive field size in the network, which controls the size of the extracted features as well as the number of layers in the network, was varied and analyzed to find its effect on generalization. In a continuation study, a comparison was made between learnt (Hebbian learning) and hard coded feature detectors. In the last study, attention focus was computed using interaction between bottom-up and top-down activation flow with the aim to handle multiple objects in the visual scene. On the basis of the results and analysis of our simulations we have found that the generalization performance of the bidirectional hierarchical network improves with the addition of a small amount of Hebbian learning to an otherwise error-driven learning. We also conclude that the optimal size of the receptive fields in our network depends on the object of interest in the image. Moreover, each receptive field must contain some part of the object in the input image. We have also found that networks using hard coded feature extraction perform better than the networks that use Hebbian learning for developing feature detectors. In the last study, we have successfully demonstrated the emergence of visual attention within an interactive network that handles more than one object in the input field. Our simulations demonstrate how bidirectional interactivity directs attention focus towards the required object by using both bottom-up and top-down effects.

In general, the findings of this thesis will increase understanding about the working of biologically-inspired interactive networks. Specifically, the studied effects of the structural and algorithmic parameters that are critical for the generalization property will help develop these and similar networks and lead to improved performance on object recognition tasks. The results from the attention simulations can be used to increase the ability of networks to deal with multiple objects in an efficient and effective manner.

Visual pattern recognition is a complex problem, and it has proven difficult to achieve satisfactorily instandard three-layer feed-forward artificial neural networks. For this reason, an increasing number ofresearchers are using networks whose architecture resembles the human visual system. These biologicallybasednetworks are bidirectionally connected, use receptive fields, and have a hierarchical structure, withthe input layer being the largest layer, and consecutive layers getting increasingly smaller. These networksare large and complex, and therefore run a risk of getting overfitted during learning, especially if smalltraining sets are used, and if the input patterns are noisy. Many data sets, such as, for example, handwrittencharacters, are intrinsically noisy. The problem of overfitting is aggravated by the tendency of error-drivenlearning in large networks to treat all variations in the noisy input as significant. However, there is one wayto balance off this tendency to overfit, and that is to use a mixture of learning algorithms. In this study, weran systematic tests on handwritten character recognition, where we compared generalization performanceusing a mixture of Hebbian learning and error-driven learning with generalization performance using pureerror-driven learning. Our results indicate that injecting even a small amount of Hebbian learning, 0.01 %,significantly improves the generalization performance of the network.

Biologically inspired hierarchical networks for image processing are based on parallel feature extraction across the image using feature detectors that have a limited Receptive Field (RF). It is, however, unclear how large these receptive fields should be. To study this, we ran systematic tests of various receptive field sizes using the same hierarchical network. After 40 epochs of training, we tested the network both by using similar but novel images of the same tropical cyclone that was used for training, and by using dissimilar images, depicting different cyclones. The results indicate that correct RF size is important for generalization in hierarchical networks, and that RF size should be chosen so that all RFs at least partially cover meaningful parts of the input image.