CN-116188842-B - Logistics illegal operation detection method in strong noise environment based on light-weight countermeasure enhancement

Abstract

The invention provides a detection method for logistics illegal operation in a strong noise environment based on light-weight countermeasure enhancement. According to the method, YOLOv is taken as a basic frame, a lightweight GhostC module is provided, a lightweight countermeasure module LAconv module is provided by applying the idea of countermeasure learning, a C3 module in the original structure is modified to be a lightweight GhostC module, a Conv module is modified to be a LAconv module, and positioning loss is modified to be CIOU loss. Finally, experiments prove that the method has excellent detection effect on the flow violation operation under the complex and strong noise background, and compared with YOLOv detection average precision average value, the method improves 1.69%, reduces 45.14% of model parameter and improves 2.46% of detection speed. The method provided by the invention has the characteristics of low parameter, high detection speed and high precision, has certain advancement and practicability for detecting the logistics illegal operation under the complex and strong noise background, and fully meets the detection requirement of the logistics illegal operation.

Inventors

• ZHANG YUAN
• ZHU LEI
• Qin Fabo

Assignees

• 北京印刷学院

Dates

Publication Date

20260505

Application Date

20221229

Claims (2)

1. 1. A method for detecting logistics illegal operation in a strong noise environment based on light-weight countermeasure enhancement is characterized in that, The method comprises the following steps: (1) Collecting three types of express abnormal operation behavior pictures of kicking, throwing and treading, taking the abnormal operation behavior pictures as an input feature picture X epsilon R Cin×H×W , wherein Cin is the number of channels, H is high, and W is wide, reducing the number of channels by the input feature picture through a function f ' conv by a scaling factor s to remove redundant channels to obtain an output feature picture Y': wherein x is input, A is weight matrix, and B is bias matrix; obtaining a brand new feature map through linear operation Where Y i is the ith feature map in Y', function The j-th linear operation is meant, Y ij is an element of the output feature set, and C out is the number of matrix dimensions; (2) Given an input feature diagram X epsilon R Cin×H×W , by global average pooling of the high H and the wide W of the compressed feature matrix, vectors u= [ u 1 ,u 2 ,u 3 ,…u Cin ],FC 1 () and FC 2 () output as channel dimensions are linear transformation functions, and the calculation formulas of the full connection layer and the activation function are as follows Wherein, sigma 1 is a ReLU activation function, sigma 2 is a Sigmoid activation function, and the calculation mode is that Modifying Conv modules in backbones and Neck into LAconv modules, taking linear transformation in Ghostconv modules as a generator G, taking a channel attention module as a discriminator D, generating a false characteristic diagram similar to a real characteristic diagram by using the linear transformation of the generator G in a lightweight contrast module, splicing the false characteristic diagram with the real characteristic diagram of a reduced channel, inputting the false characteristic diagram into the channel attention module as the discriminator to discriminate the authenticity of generated data, and forming a contrast relationship between the false characteristic diagram and the real characteristic diagram; (3) Setting a loss function by considering the relation between the transverse-longitudinal ratio and the predicted frame and the real frame Wherein IOU is an intersection ratio loss function, ρ (b, b gt ) is the Euclidean distance of the central points of the minimized real frame and the predicted frame, w gt 、h gt represents the width and the height of the target frame respectively, and w and h represent the width and the height of the target frame respectively; (4) According to the step (3), each picture is input into a network to obtain a predicted label, training loss is obtained by comparing the predicted label with a real label, parameters which can be optimized in the network are optimized by loss feedback, the steps are repeated until the set training round number is finished, finally, the prediction is completed, the image is transmitted into the network, and loss feedback is not performed.
2. 2. A method for detecting logistical violations in a high noise environment based on lightweight countermeasure enhancement as claimed in claim 1, further comprising the steps of: (5) Model evaluation index calculation Calculating the proportion P of the predicted frames with the actual violations and predicted to have the unusual behavior violations to all the predicted frames Calculating the ratio R of a predicted frame with real illegal operation to a manual annotation frame, wherein the predicted frame is predicted to have illegal operation Wherein TP represents the number of samples of which the positive class is identified as the positive class, FP represents the number of samples of which the positive class is identified as the negative class, and FN represents the number of samples of which the negative class is identified as the negative class; calculating the average mAP of three average accuracies of kicking, throwing and treading Wherein AP is the area surrounded by the curves formed by P and R, and n is the number of three behaviors of kicking, throwing and treading.

Description

Logistics illegal operation detection method in strong noise environment based on light-weight countermeasure enhancement Technical Field The invention relates to the field of intelligent recognition of express logistics, in particular to a method for detecting logistics illegal operation in a light-weight countermeasures and enhanced strong noise environment. Background With the development of logistics industry, the problems of damage, breakage and the like of the package in logistics operation are serious, and the direct reason for the phenomenon is illegal operation in logistics operation. The illegal operation refers to the action of performing rough operation on the package by an operator in the whole logistics links of package collection, transfer, storage, distribution and the like, wherein the illegal operation is particularly prominent in the aspect of express delivery. Logistics violation operations are very common in the express industry, wherein at least 16.1% of express items are damaged to different degrees. The user complaint condition notice of the postal service of 2021 month 12 shows that express damage accounts for 21.8% of the total complaint amount in the main problem of complaint of express service. At present, the identification of the logistic illegal operation mainly comprises two technical routes, namely a first identification method based on a sensor cluster and a second identification method based on videos and images. The logistic illegal operation identification method based on the sensor cluster is characterized in that sensors such as inertia and vibration are built in the package, the illegal operation is analyzed and processed through development of special hardware and a cloud platform, the express violence sorting identification method based on acceleration distribution characteristics and provided by Ding Ao is representative, the illegal operation of the package is effectively identified and classified, and the method has important significance for monitoring the whole flow of the package. However, when the method is used for identification, the fixed installation mode of the special detection terminal is relatively complicated and occupies the packaging volume, and meanwhile, the special detection terminal is generally powered by a lithium battery, so that the safety of transportation modes such as aviation and the like is adversely affected. Therefore, there is a certain disadvantage in detecting the operation of a logistic violation based on a sensor cluster. In recent years, methods for detecting target behaviors based on videos and pictures are rapidly developed, and are widely used in various fields of agriculture, medical treatment, industrial manufacturing, and the like. Intensive research has been conducted by the present scholars with respect to the identification of the operation of violations of logistics. Wu Pengbo and the like, the express violence sorting detection system based on the LSTM+Attention and MobileSSD model can conduct behavior recognition through gesture data. Shang Shuling and the like effectively extract the collected image behavior characteristics of the logistics sorting by utilizing a wavelet packet analysis method, and provide a basis for identifying the logistics illegal operation. Deng Xiuqin et al also propose a computer vision-based violent sorting behavior identification method for logistics violation operations. However, the method has lower detection effect and stability and high model training and detection cost. Therefore, a light and stable detection method has important theoretical and practical significance. Disclosure of Invention Aiming at the problem that the logistics illegal operation is difficult to effectively identify under the complex and strong noise background, the invention provides a lightweight anti-reinforced logistics illegal operation detection method. Aiming at the problem of logistics violation operation, the invention provides a light-weight countermeasure detection method for a complex strong noise background by taking YOLOv network as a main frame, which solves the problems of high detection cost and poor detection precision of logistics violation operation and effectively reduces the workload of manual intervention. The invention adopts the following technical scheme: a detection method for logistics illegal operation in a strong noise environment based on light-weight countermeasure enhancement comprises the following steps: (1) Collecting three types of express abnormal operation behavior pictures of kicking, throwing and treading, taking the abnormal operation behavior pictures as an input feature picture X epsilon R Cin×H×W, wherein Cin is the number of channels, H is high, and W is wide, reducing the number of channels by the input feature picture through a function f 'conv by a scaling factor s to remove redundant channels to obtain an output feature picture Y': Y'=X×f'conv f'conv=∑Ax+B wherein