Deep learning algorithm based on multi-task and nearby information fusion for object detection

Abstract

The invention discloses a multi-task and proximity information fusion depth learning algorithm based on object detection, which comprises the steps: pictures are input, image features by using a convolution neural network are extracted, and target candidate frames are generated; using the image features, the image is passed through the region candidate network to extract the target prediction frame; feature extraction and feature pooling of the target prediction frame, then the border regression, direction prediction, target detection and classification are performed to get the preliminary detection results; the preliminary detection result is fused with the target candidate frame and enters the ROI pooling layer and passes through the second full connection layer to obtain the final detection result; the classification of target detection is to redefine the confidence score of a target prediction box by using the information relationship between the target prediction box and its neighboring other target prediction boxes. The algorithm adopts multi-task output mode. The invention not only improves the speed of the target detection, but also ensures the accuracy of the target detection and achieves the requirement of the real-time target detection.

Description

technical field [0001] The invention relates to the field of image information processing, in particular to a multi-task based target detection and a deep learning algorithm for fusion of adjacent information. Background technique [0002] At present, target detection has always been a basic problem in the application of visual computing, which is applied in traffic monitoring, intelligent driving and other fields. In real conditions, on the one hand, due to the diversity of targets such as vehicles, pedestrians, numbers, railings, etc. on the road, there are many subcategories of targets such as buses, cars, trucks, bicycles, etc. on the other hand. The target has multiple angles, different occlusion situations and the local size of the target display, which brings great difficulty to target detection. Target detection is still a very challenging field, and in order to achieve target detection, recognition, and tracking in real time, there are quite high requirements for d...

AI Technical Summary

Problems solved by technology

[0003] Starting from RCNN, namely Regions with CNN features, the convolutional neural network was introduced into the field of target detection, which greatly improved the effect of target detection; subsequently, SppNET, Fast-RCNN and Faster-RCNN were proposed to further accelerate the convolutional neural network in target detection. field of application, but at the same time there is a contradiction between accuracy and recognition speed, the recognition speed is reduced due to repeated feature extraction and calculation, and a large storage space is required

In addition, there is also a structure of YOLO, which is You only look once, which has a fast recognition speed, but at the expense of a certain accuracy rate.

Images