Fine-grained classification of automobile front face modeling based on Gestalt psychology

In this paper, we propose a fine-grained classification method for automobile front face modeling images based on Gestalt psychology. This method divides pixels into features of visual regions through convolutional neural network, divides automobile front face images into parts, and conducts fine-grained classification based on the overall modeling of parts. A more objective method of fine granularity classification of automobile front face image is explored. A fine-grained classification and recognition model of automobile front face modeling based on Gestalt psychology is proposed in this work. Firstly, unclassified input car front face images are filtered through part detection, part segmentation, and regularization processing by combining the image classification training sets of car front face shapes. Secondly, to facilitate weakly supervised learning for each part, we establish recognition models using the simple a priori of U-shaped distribution for individual parts of car images and train the net using image-level object labels on the ResNet-101 network framework. Attention mechanism is then reused for aggregate features to output classification vectors. Finally, recognition accuracy of 89.9% is reached on the Comprehensive Cars (CompCars) dataset. Compared with other CNN methods, the results confirm that U-shaped distribution combined with parts in the exploration image has a higher recognition rate. Moreover, model interpretability can be achieved by dividing images and recognizing the contribution of each part in the classification.