Kotwaliwale, NachiketSabat, Mousumi2022-04-302022-04-302021https://krishikosh.egranth.ac.in/handle/1/5810184035T-10653Drying, is one of the most common primary shelf-life enhancing technique. However, the process causes many sensorial, physical and mechanical changes in the food products, presently being evaluated by human observers or offline chemical methods for quality detection. However, these measurements are tedious, time consuming and subjective in nature. Assessment of the qualitative changes taking place during drying using computer vision system provides better alternative for process monitoring and control since it is rapid, inexpensive, non-destructive, sensitive, and precise. The computer vision system (CVS) includes five primary steps viz., image acquisition, image processing, feature extraction, pattern recognition and decision making. Previously, researchers were using different software for image acquisition and offline image processing. Whereas with online monitoring of the drying phenomena, changes in its size, shape, colour, texture, etc. can be studied and monitored as they are complexly dependent on each other as well as on the drying parameters. A laboratory scale dryer was developed and enable with sensors, controls and air distribution systems to ensure uniform distribution of temperature and air flow throughout the chamber. To visualize the colour, texture and morphological changes during drying progression, the dryer was equipped with an image acquiring system. In the current study experiments were conducted on drying of potato slices. Drying was monitored on real-time basis during at 4 levels of the drying air temperature (T) i.e., 45, 50, 55 and 60 °C, 3 levels of potato slice thickness (L) i.e., 1, 2 and 3 mm and 1.5 m/s of air velocity. The sample weight data was recorded till constant weight was achieved. An image acquisition system consisted of four principal components viz., camera, illumination, hardware and software. The images were processed using the algorithms written in MATLAB® platform. Image features were extracted to obtain the various quality related parameters (colour, image texture and product morphology) about the object of interest. Broadly image features can be classified into two main categories viz., external and internal. The internal image features include the chromatic and image textural features whereas the external features include the morphology of the product. 93 For studying the time series data of the image properties at different temperatures and slice thicknesses; recurrent neural network (RNN) was used that predicts the future while using the past data. Since three categories of image properties, viz., morphological, chromatic and textural properties, were studied; initially the attributes from each property were used as an input to the network to observe the predictability of the output (moisture content) and later the combinations of two and all three properties were used. Additionally, different number of nodes, number of hidden layers, transformation function and training algorithm used were determined, on hit and trial basis, since it depended on the complexity of the problem to be investigated. Using the best RNN (that has highest R2 value and lowest MSE as well as MAPE value) the algorithm was developed with the intent of creating a graphical user interface (GUI). The developed GUI helped in simplifying communication of information between the system and operator, being user friendly. To check the efficiency of the developed algorithm, the moisture content of the product displayed during drying was compared with the moisture content obtained from the sample weight (manually recorded). All morphological parameters considered in this study, except eccentricity and roundness, decreased smoothly till normalised moisture content (NMC) is equal to 0.25 and then declined rapidly. The roundness and eccentricity values increased after NMC0.97). In RNN 3, the network consisting of three hidden layers with 15-5-10 number of nodes in first, second and third hidden layers, respectively was selected as optimal topology to predict the moisture content of the potato slices during drying. In order to evaluate the performance of the developed algorithm on real time basis, the drying data were taken manually and compared with the moisture content predicted by the developed algorithm at 47, 54 and 58 °C for slices of thickness 2, 1 and 3 mm, respectively. The statistical parameters R2 (0.9981, 0.9991 and 0.9957), MSE (2.77 × 10-4, 1.62 × 10-4 and 8.88 × 10-4) and MAPE (0.0708, 0.3668 and 0.2242) indicated that the developed network as well as the algorithm perform efficiently.EnglishThesis