Small target detection processing method for railway cargo loading state image

Abstract

The invention discloses a small target detection processing method for a railway cargo loading state image. The method comprises the following steps: cutting a large image by using mutually overlappedoverlay; introducing a small target annotation size variable in the process of cutting the large graph; cutting the large graph to obtain a small graph in which all original feature values are reserved; carrying out image recognition by using the maximum equal-ratio neuron number of the small image pixels; recording the change proportion and the change quantity of the weight parameters in the neural network training process; training the training data set by using a neural network structure to obtain a network weight, and completing network training; segmenting the large graph to obtain smalltarget recognition input data; carrying out small target recognition on the input data by using the network weight obtained in S6. According to the invention, image recognition is carried out on small images of all original feature values, so that small target features are prevented from being lost; and training the training data set by using a neural network structure to obtain a network weight,and carrying out small target recognition on the input data.

Description

technical field [0001] The invention relates to the technical field of railway cargo image detection, in particular to a small target detection and processing method for railway cargo loading status images. Background technique [0002] The cargo inspection station is an important node of the railway transportation network. Cargo inspection stations are used to detect the loading status of the cargo, such as deviation of the center of gravity, foreign objects on the side of the car, foreign objects on the roof, opening of the car cover, damage to the tarpaulin, opening of the box car door, opening of the middle door of the gondola car, etc. Issues that affect transportation safety, such as empty vehicles or not empty, are an important link to ensure the safe transportation of goods. At present, the inspection of cargo status at cargo inspection stations mainly relies on manual map judgment, manual inspection, etc., and manual inspection is easily affected by physiological and...

AI Technical Summary

Problems solved by technology

[0002] The cargo inspection station is an important node of the railway transportation network. Cargo inspection stations are used to detect the loading status of the cargo, such as deviation of the center of gravity, foreign objects on the side of the car, foreign objects on the roof, opening of the car cover, damage to the tarpaulin, opening of the box car door, opening of the middle door of the gondola car, etc. Issues that affect transportation safety, such as empty vehicles or not empty, are an important link to ensure the safe transportation of goods. At present, the inspection of cargo status at cargo inspection stations mainly relies on manual map judgment, manual inspection, etc., and manual inspection is easily affected by physiological and psychological factors. Influence, high strength, low efficiency, especially in the small object problem items under the large background image, there is obvious difficulty in discrimination

[0003] The small target object of the railway cargo loading status image refers to the object object whose size accounts for 0.01% to 0.2% of the large background railway cargo loading status image, that is, the pixel range of the object object is 30×60 pixels to 160×200 pixels ;Small target detection mainly exists in problem points such as foreign objects on the side of the car, foreign objects on the roof, open hood, and damaged tarpaulin. Due to the large background image and small problem points, it is difficult to identify with the naked eye, and there are cases of missed and false detections. The land has increased the risk and hidden dangers of railway freight safety management

[0004] Image recognition algorithms represented by Yolo, SVM, etc., are prone to lose small target features when images of medium size, such as 1024×768 pixels, are reduced and processed; the maximum pooling algorithm establishes a large-scale negative sample image set to train sample images The size is too small, and the algorithm process is difficult to complete large image recognition in a short time; in large images, such as 2048×8000 pixels, after shrinking, the small target basically becomes a single-digit pixel width, the signal is lost, and it cannot be recognized, so the above algorithm cannot meet the requirements of large images. Requirements for small target recognition in background images; public detection data show that Google, Microsoft, and IBM test data sets, small target detection must use small target special algorithms, and calculations based on general models cannot achieve small target detection

Images