End-to-end weak supervision target detection method based on frame regression of deep learning

Abstract

The invention discloses an end-to-end weak supervision target detection method based on frame regression of deep learning. According to the method, in a weak supervision convolutional neural network,a frame subjected to convolution layer and selective search is subjected to pyramid pooling layer and two full connection layers through a feature map, then a feature vector of a prediction frame is output, and then a full connection layer and a softmax layer on the category are connected; and finally, a prediction score corresponding to each object class in the selective search box is output. A box with the highest score of each class is selected as a pseudo annotation frame of the class; and the object frame predicted by the weak supervision model is regressed by using the frame with the highest score detected by each category as a pseudo labeling frame to generate a regression loss function, and a new loss function supervision weak supervision detection model is formed by the classification of the regression loss function and the weak supervision model and the positioning loss function. According to the invention, the detection time is reduced, and the target detection efficiency isimproved.

Description

technical field [0001] The invention relates to the field of deep learning target detection and artificial intelligence, in particular to an end-to-end weakly supervised target detection method based on deep learning frame regression. Background technique [0002] Object detection is a basic problem based on computer vision, and it plays a huge role in medical diagnosis, security monitoring and unmanned driving. At the same time, with the rapid development of machine learning, the hardware field has also been greatly improved. In recent years, many excellent target detection models have emerged. However, the current target detection models with good performance are basically fully supervised target detection with finely labeled images containing bounding boxes as the conditions for target detection supervision. This requires a lot of manpower and material resources to label the pictures with fine borders. However, the annotation information of the border also has many shor...

AI Technical Summary

Problems solved by technology

The two-step convolutional neural network model increases the time of target detection and reduces the efficiency of target detection

Images