TOP  >  Thesis/Dissertation  >  Ingredient Recognition from Load Feature at Cutting Action

Ingredient Recognition from Load Feature at Cutting Action

Because cooking is complicated work on daily basis, a supporting system is expected that gives advices to chefs with considering their current situation in cooking. To understand the situation of a chef, it is desired that the system recognizes ingredients automatically based on sensor information. In cooking, ingredients get culinary operations, and they almost lose their unique characteristics. Because of such affection to the characteristics, it is difficult to recognize ingredients after the operations. Generally, the first operation to ingredient is a cutting action.Therefore, we can extract the raw characteristic of an ingredient as a feature at the timing of cutting action.A cutting action is done by many chops. And during a chop, the collision with an ingredient and a kitchen knife changes the load to the chopping board. The load during a chop may consist of two different kinds of forces: by pushing the kitchen knife to the chopping board, and by supporting the ingredient by hand. Especially, the force pushing the knife is affected by the physicality of ingredient when the knife passes inside of the ingredient. Using this effect, we propose a method for recognizing ingredients from load feature in this paper. To recognize ingredients from the chops in real cooking scene, it needs firstly to measure load, secondly to detect chop, and thirdly to extract features. To measure the load on a chopping board, we developed the load sensing board using load sensors. The proposed method detects a chop by finding sharp falling edge of load. The sharp falling edge corresponds to the end of each chop because the both two kinds of pressures peg out simultaneously at the end of a chop. Next, the method clips chop segments from detected chops, and extracts Max, Duration, Impulse, Peak Position, and Kurtosis as a feature. Finally, the method realizes ingredient recognition by applying SVM to the extracted features. To evaluate this method, we tested the method with 16 kinds of ingredients. In this evaluation, we used observed data including any possible actions on chopping board, targeting a real cooking scene. As a result, the proposed method achieved a precision rate of 98.1% and a recall rate of 99.8% in chop detection, and a recognition rate of 67.4% in ingredient recognition.