Development of methodologies for segmentation recognition and analysis of indian food items from images

Abstract

Recognition of Indian food can be assumed to be a fine-grained visual task owing to recognition property of various food classes. It is therefore important to provide an optimized approach to segmentation and classification for different applications based on food recognition. Food computation mainly utilizes a computer science approach which needs food data from various data outlets like real-time images, social flat-forms, food journaling, food datasets etc, for different modalities. In order to consider Indian food images for a number of applications we need a proper analysis of food images with state-of-art-techniques. The appropriate segmentation and classification methods are required to forecast the relevant and upgraded analysis. As accurate segmentation lead to proper recognition and identification, in essence we have considered segmentation of food items from images. Considering the basic convolution neural network (CNN) model, there are edge and shape constraints that influence the outcome of segmentation on the edge side. Approaches that can solve the problem of edges need to be developed; an edge-adaptive Deep CNN approach is proposed in this work to provide an efficient segmentation. However, edge adaptive (EA) is directly associated with DCNN, which is advantageous both in the training process and in the relationship between spatial data. The output obtained from DCNNs is feed forwarded to EA approach to get more refined output. The EA- DCNN model includes a rectified linear unit (RLU), convolution, and pooling to optimize food image segmentation. As we have solved the problem of food segmentation with EA-DCNN, we also have difficulty in classifying food, which has been an important area for various types of applications. newlineFood analysis is the primary component of health-related applications and is needed in our day-t- day life. Therefore, we consider Google Inception-V3 based CNN method for the classification of food image. It has the proficiency to directly predict the score function from image pixels, input layer to produce the tensor outputs and convolution layer is used for self- learning kernel through back-propagation. In this method, feature extraction and Max-Pooling is considered with multiple layers, and outputs are obtained using softmax functionality. The proposed implementation tests 92.89% accuracy by considering some data from yummly dataset and by own prepared dataset. Consequently, it is seen that some more improvement is needed in food image classification. We therefore consider the segmented feature of EA-DCNN and newline newlineconcatenated it with the feature of our custom Inception-V3 to provide an optimized classification. It enhances the capacity of important features for further classification process. In extension we have considered south Indian food classes, with our own collected food image dataset and got 96.27% accuracy. The obtained accuracy for the considered dataset is very well in comparison with our foregoing method and state-of-the-art techniques. With classification, analysis of food image for calorie, ingredient and properly baked or not are also considered in our work. newline newline