Fruit image recognition method based on multi-task learning

Abstract

The invention discloses a fruit image recognition method based on multi-task learning, which comprises the following steps of: firstly, constructing a recognition model by adopting a symmetrical encoder-decoder network structure, randomly extracting a batch of images from a fruit image data set, carrying out gray scale transformation preprocessing on each image, and inputting a gray image of the obtained image into the recognition model; completing the construction of a classification task, and then completing the construction of a coloring task by using the colored image; calculating a cross entropy loss function of an image category output value and an image category true value, calculating a mean square error loss function of a colored image and an original input image, and optimizing parameters in a network by using a weighted sum between the two functions so as to complete multi-task learning; and finally, carrying out online testing on the fruit image by utilizing the trained identification model. By adopting an encoder-decoder network structure, the classification model can extract texture and color features of fruit images at the same time, and the accuracy of fruit classification and recognition is improved.

Description

technical field [0001] The invention relates to a fruit image recognition method based on multi-task learning, and belongs to the field of computer vision. Background technique [0002] The research on automatic classification and identification of fruits has important application value in social life and production. For example, in the process of fruit sales, the automatic identification of fruits can bring convenience to people's purchase, thereby promoting the development of unmanned supermarkets; In the production process, the automatic identification of fruits can greatly improve the efficiency of automatic fruit picking, thereby promoting the development of smart agriculture; in the process of daily life, the automatic identification of fruits can help to analyze the nutritional components of fruits and help people formulate a reasonable diet. collocation, thereby helping to improve people's quality of life. [0003] With the rapid development of multimedia technology...

AI Technical Summary

Problems solved by technology

For example, Lu Benxue et al. (Lv Benxue, Zhao Jun, Tian Chunyang, Sun Kesi. A method of fruit recognition based on deep learning. Application number: 202010181484.X) disclosed a method of fruit recognition based on deep learning. The convolutional neural network of the accumulation layer and the pooling layer extracts the features of the fruit image and uses the Softmax classifier for classification and recognition; Wang Jiahong et al. A fruit recognition method and device based on deep learning. The method uses the Alexnet network pre-trained on ImageNet and fine-tunes fruit pictures to realize fruit classification and recognition. However, the convolutional neural network used in the above method only focuses on fruit recognition. The texture feature of the image ignores the color feature information of the fruit image. In order to overcome the above technical problems, it is necessary to disclose a fruit image recognition technology based on multi-task learning, so that the classification model can simultaneously extract the texture and color features of the fruit image, and improve the quality of the fruit image. The correct rate of classification recognition

Images