A Research To Develop An Invariant Recognition And Classification Method Based On Biologically Inspired Color Vision Model

Human Vision Systems has always been an interesting topic for machine vision researchers to mimic the kind of intelligence and visual perception system possessed by humans and other mammals.Since the human visual perception system is still par better than most advance machine vision algorithms,therefore it has been a great interest to study the physiology and anatomy of the human eye and visual cortical processing areas and then try to mimic them in machine vision algorithms in order to achieve human like intelligence and vision.The main objective of this thesis is to propose a machine vision algorithm structure based on the neurological and physiological evidences found by several neuroscientists recently and then apply the proposed method for the industrial process control and automation.In this study,a new model(Color-HMAX-SVM)is proposed for color machine vision based on the hierarchal architecture inside human brain(HMAX model)to represent the image pattern(or texture)and color features invariant to the scale,position,orientation or other geometric or illumination variances,and to recognize and classify the recognized patterns according to their texture(or pattern)and color to their respective class.Since the features of an object are unique in general,in terms of texture or pattern as well as color,the developed algorithm must be capable to capture both the pattern and color features appropriately in order to make the recognition and classifications effectively.Since biological vision systems are quite robust in representing the visual information invariantly,therefore our work was inspired by the established facts about the primary visual cortex.The developed algorithm can be divided into three modules namely the color descriptor,the shape descriptor and the feature classifier.We developed a pattern descriptor by creating a color variant of the H-MAX model.This was accomplished by developing a three opponent channel color descriptor based on the formation of three opponent color channels(just like in humans namely,red-green,blue-yellow and black-white)from the input color image to extract and represent the color features from the image and then integrating it with the conventional H-MAX model.The H-MAX model is based on a hierarchal architecture with alternating convolution and max pooling layers and is capable of capturing features relating to the pattern,and opponent channel color descriptor captures the color information.A feature classifier based on the support vector machine(SVM)was developed and integrated with the feature extraction module to perform the feature matching as the final stage of our model.The model was tested on a custom-built fabric pattern dataset and was found to be considerably efficient in recognizing different patterns and colors.Furthermore,the developed model discussed in previous paragraph(Color-HMAX-SVM)is improved by replacing the SVM classifier in the Color-HMAX-SVM model with the deep ELM based feature classifier.Although the previously proposed model was biologically inspired and quite efficient,but we realized that the classifier could be improved in order to improve the classification accuracy of the model proposed in the previous study.Several researchers also have a difference of opinion in the biological plausibility of the support vector machine(SVM)classifiers but the extreme learning machines are proved to be biologically plausible and also a better alternative as compared to other existing classifiers in terms of simplicity,efficiency and speed.Recently,during this decade extreme learning machines have emerged as a powerful alternate as a feature classifier,the extreme learning machine based classifiers have proved to be easier to train,faster,less computation extension as compared to the support vector machine counterparts.This had led to inspire us to propose improved model for the pattern recognition and classification to enhance the previously proposed model efficiency.Therefore,we built another model for the color machine vision,also inspired by the established facts about the human visual system.We developed a deep extreme learning machine network as a feature classifier and integrated it with the previously developed pattern feature descriptor.The developed algorithm(Color-HMAX-DELM)was then applied to the custom-built fabric pattern dataset to test the model efficiency.The achieved results were then compared to other existing models as well as the previously proposed model,and it was found that the achieved results are superior to other existing models and the previously proposed model(Color-HMAX-SVM)Finally,the proposed models are applied to industrial automation.We realized that the current trend of most industries is the automated process control in order to reduce labor cost,improve efficiency and production speed.The quality control has always been an important factor for any company,whether it is a textile manufacturer or any other consumer goods manufacturer.Firstly,the proposed model is applied to the problem of biscuit baking level and texture inspection to quantitatively test the recognition accuracy of the model for similar patterns but different color levels.The baking level of the baked food is an important feature to assess the quality of the baked food and the baking level is judged subjectively by the color appearance of the food,therefore by performing the baking level test of the baked product we verified the color feature recognition effectiveness of our developed algorithm.The developed algorithm was tested over custom-built biscuit dataset images,and the achieved results were compared with the existing machine vision algorithms.The comparison of the results was found to be satisfactory.Secondly,the proposed model is applied to the problem of recognition and classification of traffic signs under varying environmental conditions to test the invariance capability of the algorithm using color information.The key problem in the development of traffic sign recognition system is to develop an intelligent vision system capable of recognizing traffic signs and other objects in varying environmental conditions such as changes in illumination.Therefore,we applied our algorithm to this problem to test the recognition accuracy of our algorithm under varying environmental conditions for the recognition of traffic signs invariant to the position,orientation,occlusion and illumination.The experimental results showed that the developed algorithm could use the color information effectively and was capable of recognizing and classifying the relevant traffic signs correctly regardless of the variance,in most cases.The developed algorithm was found to be robust and not miss the target in most cases.Thirdly,we applied the developed biologically inspired color machine vision algorithm for the complex real industrial problem,invariant recognition and classification of the textile fabric according to the weave pattern and yarn color.The macro-level visual inspection of the fabric weave pattern and yarn color is an important industrial application as it is based on the textural as well as the color assessment of the textile fabric.We applied the developed algorithm to this important industrial application by testing it upon a custom-built fabric image dataset and then compared it with other existing methods.The experimental results and the comparison showed that the developed algorithm has an impressive recognition capability from color images under varying conditions and that the fabric weave patterns and yarn color were recognized and classified to the respective class category quite accurately.The major work in this thesis was to imitate human vision and human intelligence to provide the machines with the vision and intelligence capabilities same as humans.We developed two kinds of algorithms and tested them over pattern recognition and classification,as well as for different applications,to find the possibility and capability of the biologically inspired artificial vision and artificial intelligent system to replace the human labor in terms of cost,speed and efficiency.The results were found to be quite impressive and the comparison with other existing models proved to be satisfactory.Since the neuroscientists are always discovering about the processing mechanisms in the primary visual cortex,therefore there is always a room for improvement in the biologically inspired algorithms for artificial vision and intelligence based on the latest discoveries about human vision system.The results suggest that further advancement in the area of biological vision system may lead to further improvement of the performance of these biologically inspired machine vision algorithms and their applications for industrial automation.