CN-122020200-A - Multi-dimensional strawberry size distribution-based web rule dynamic matching method

CN122020200ACN 122020200 ACN122020200 ACN 122020200ACN-122020200-A

Abstract

The invention discloses a net-cover rule dynamic matching method based on multi-dimensional strawberry size distribution, which comprises the steps of A1, collecting strawberry double-view images and strawberry weight data, respectively preprocessing, A2, respectively extracting initial morphological characteristics, then carrying out double-view characteristic fusion to obtain double-view fusion characteristics, A3, inputting the double-view fusion characteristics into a fruit-shaped contour enhancement network to obtain contour enhancement characteristics, A4, extracting strawberry three-dimensional morphological characteristics, A5, extracting strawberry weight characteristics, then carrying out heterogeneous characteristic fusion in combination with the strawberry three-dimensional morphological characteristics to obtain strawberry fusion characteristics, A6, extracting initial net-cover rule characteristics and final net-cover rule characteristics, A7, extracting target net-cover rule labels according to the final net-cover rule characteristics, and finally completing net-cover packaging operation of strawberries. The invention can solve the problem of poor packaging protection effect caused by the difficulty in adapting to the shape difference of the strawberries in the traditional packaging method.

Inventors

ZHU XIANYU
Pang Zonghao

Assignees

广西豪融农业科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260129

Claims (8)

1. A web rule dynamic matching method based on multi-dimensional strawberry size distribution is characterized by comprising the following steps: a1, acquiring strawberry double-view images and strawberry weight data, and respectively preprocessing to obtain preprocessed strawberry double-view images and preprocessed strawberry weight data; a2, respectively extracting initial morphological characteristics for each view according to the preprocessed strawberry double-view image, and then carrying out double-view characteristic fusion to obtain double-view fusion characteristics; A3, inputting the double-view fusion features into a fruit-shaped contour enhancement network, and enhancing and optimizing the contour features to obtain contour enhancement features; a4, extracting three-dimensional morphological characteristics of the strawberries according to the contour enhancement characteristics; A5, extracting weight characteristics of the strawberries according to the pretreated weight data of the strawberries, and combining the three-dimensional morphological characteristics of the strawberries to perform heterogeneous characteristic fusion to obtain strawberry fusion characteristics; A6, extracting initial net nest specification characteristics and final net nest specification characteristics according to the strawberry fusion characteristics; A7, according to the final mesh specification characteristics, finishing characteristic mapping through a multi-layer perceptron to obtain mesh specification classification characteristics, extracting target mesh specification labels through a Softmax function and a Argmax function, and finally finishing mesh packing operation of strawberries.
2. The web rule dynamic matching method based on multi-dimensional strawberry size distribution according to claim 1, wherein the A1 step comprises: a11, acquiring a strawberry double-view image through an industrial camera, wherein the data type is a color pixel matrix and comprises two groups of sub-view data of a strawberry top view and a strawberry side view; A12, preprocessing the strawberry double-view image data by adopting a Gaussian filter denoising and bilinear interpolation resolution unification method to obtain a preprocessed strawberry double-view image; A13, preprocessing the strawberry weight data by adopting a triple standard deviation constant value eliminating and maximum and minimum value normalizing method to obtain preprocessed strawberry weight data.
3. The web rule dynamic matching method based on multi-dimensional strawberry size distribution according to claim 1, wherein the A2 step comprises: a21, respectively extracting initial morphological characteristics for each view angle according to the preprocessed strawberry double-view angle image; A22, carrying out double-view feature fusion according to the initial morphological features of the two views to obtain double-view fusion features; The extraction process of the initial morphological feature comprises the steps of carrying out convolution layer, batch normalization and LeakyReLU function processing on a single-view image to obtain an original image feature, constructing a view feature calibration function, carrying out batch normalization, 1X 1 convolution layer, adding a calibration bias term and Sigmoid function processing on the original image feature, and carrying out Hadamard product operation on the operation result of the original image feature and the view feature calibration function to obtain the initial morphological feature; the extraction process of the double-view fusion feature comprises the steps of splicing initial morphological features of two views, obtaining view gating weights of all views through layer normalization, a multi-layer perceptron, gating scaling factors and Sigmoid function calculation, carrying out Hadamard product on the view gating weights of all views and the corresponding initial morphological features, adding the results element by element, and carrying out GeLU function and layer normalization treatment to obtain the double-view fusion feature.
4. The web-style rule dynamic matching method based on multi-dimensional strawberry size distribution according to claim 3, wherein the fruit-style contour enhancement network in the A3 step comprises: The method comprises the steps of processing a convolution layer with a convolution kernel size of 3 multiplied by 3, adding the convolution layer with the convolution kernel size to the overall average pooled double-view fusion feature element by element to obtain a double-view fusion residual feature, splicing images of two views in a preprocessed strawberry double-view image, obtaining a fruit-shaped template matrix through convolution layer, reLU function and transposition convolution, processing the double-view fusion residual feature by a multi-layer perceptron and Sigmoid function, processing the double-view fusion residual feature and the double-view fusion residual feature by a Laplace operator, then making Hadamard product with the result of Hadamard product with the fruit-shaped template matrix, and finally adding the Hadamard product with the double-view fusion residual feature to obtain the contour enhancement feature.
5. The web rule dynamic matching method based on multi-dimensional strawberry size distribution according to claim 4, wherein the extraction process of the stereoscopic morphological features of the strawberries in the step A4 comprises the following steps: Firstly, processing the contour enhancement features through a convolution layer with the convolution kernel size of 3 multiplied by 3, simultaneously splicing view gate weights of two views, processing the two processing results through a multi-layer perceptron, and processing the two processing results through a Sigmoid function and a convolution layer with the convolution kernel size of 1 multiplied by 1 to obtain a fruit-shaped attention constraint matrix; the average value of two view angle gating weights and a fruit-shaped attention constraint matrix are adopted to make Hadamard products, 1 is added, and then Hadamard products are sequentially made with contour enhancement features and contour enhancement features processed by an attention mechanism, so that contour attention features are obtained; And the contour attention feature is sequentially processed by a convolution layer with a convolution kernel size of 3 multiplied by 3, a ReLU function and a convolution layer with a convolution kernel size of 1 multiplied by 1, and then added with the contour enhancement feature element by element to obtain the strawberry stereoscopic morphological feature.
6. The web rule dynamic matching method based on multi-dimensional strawberry size distribution according to claim 5, wherein the A5 step comprises: a51, extracting weight characteristics of the strawberries according to the pretreated weight data of the strawberries; A52, carrying out heterogeneous feature fusion according to the three-dimensional morphological features of the strawberries and the weight features of the strawberries to obtain strawberry fusion features; The calculation process of the strawberry fusion characteristic comprises the following steps: firstly, carrying out Hadamard product operation on three-dimensional morphological characteristics of strawberries and weight characteristics of strawberries, and processing an operation result by a multi-layer perceptron and a Softmax function to obtain a multi-source characteristic attention fusion matrix; Carrying out Hadamard product on the strawberry three-dimensional morphological characteristics and the multisource characteristic attention fusion matrix, and carrying out element-by-element addition on the Hadamard product and the strawberry three-dimensional morphological characteristics to obtain calibrated strawberry three-dimensional morphological characteristics; And processing the calibrated three-dimensional morphological characteristics of the strawberries through a convolution layer, processing the result obtained by multiplying the weight characteristics of the strawberries and the weight matrix of the weight characteristics of the strawberries through a multi-layer perceptron, and carrying out Hadamard product operation on the two processing results to obtain the strawberry fusion characteristics.
7. The web rule dynamic matching method based on multi-dimensional strawberry size distribution according to claim 5, wherein the extraction process of the final web rule feature in step A6 comprises: Multiplying the strawberry fusion characteristic with a mesh specification weight matrix, and processing by a ReLU function, batch normalization and a multi-layer perceptron to obtain an initial mesh specification characteristic; And processing the initial net nest specification feature through a convolution layer, processing the strawberry fusion feature through a multi-layer perceptron, carrying out Hadamard product operation on the two processing results, and adding the Hadamard product operation with the initial net nest specification feature element by element to obtain the final net nest specification feature.
8. The web rule dynamic matching method based on multi-dimensional strawberry size distribution according to claim 5, wherein the A7 step comprises: A71, according to the final mesh specification characteristics, finishing characteristic mapping through a multi-layer perceptron formed by three full-connection layers to obtain mesh specification classification characteristics; a72, according to the classifying characteristics of the net nest specifications, firstly, calculating to obtain probability distribution results corresponding to all net nest specifications through a Softmax function, and then selecting an index corresponding to a result with the maximum probability value in the probability distribution results through a Argmax function so as to determine a target net nest specification label matched with the current strawberry; And A73, inputting the target web specification label into a strawberry packaging execution system, and calling the execution system to finish the web packaging operation of the strawberry.

Description

Multi-dimensional strawberry size distribution-based web rule dynamic matching method Technical Field The invention relates to the technical field of fruit packaging, in particular to a net set rule dynamic matching method based on multidimensional strawberry size distribution. Background The strawberry is used as a fragile and damaged fruit, the package of the net sleeve is a key step for guaranteeing the strawberry from collision damage in the transportation and storage processes in the packaging link from the picking to the selling, the core is that the friction and extrusion between strawberries are reduced by utilizing the buffering and limiting effects of the net sleeve through the net sleeve matched with the strawberry in size and shape, the corrosion rate is further reduced, and the general process generally comprises the steps of strawberry harvesting, pretreatment, specification matching, net sleeve sleeving, subsequent boxing and the like. However, in practical application, because the shape difference of the strawberries is remarkable, the strawberries have regular conical shapes, oblate shapes and other special-shaped fruits, the sizes and the three-dimensional shapes of the single strawberries are different, and a plurality of pain points exist when the size adaptation of the net cover and the strawberries is realized. The traditional method for solving the adaptation problem mainly comprises the steps of adding a strawberry sorting step before mesh sleeve matching, dividing strawberries into different grades according to indexes such as fruit diameter size, fruit shape appearance and the like through manual or sorting equipment, and matching strawberries of the same grade with mesh sleeves of uniform specification, wherein in the mode, on one hand, strawberries collide with each other and are easy to cause skin damage during sorting, the later corrosion risk is increased, on the other hand, the time length of production line operation of a production place is greatly prolonged by an additional sorting process, the labor and equipment investment cost is increased, even if classification is finished, the fruit shapes of strawberries of the same grade still have obvious differences, the uniform mesh sleeve is easy to cause the problems that flat strawberries are not tightly attached to the mesh sleeve, shake and scratch is damaged in transportation, and the conical strawberries are pressed by the mesh sleeve to cause pulp damage. With the application expansion of the deep learning technology in the agricultural product processing field, at present, partial schemes try to assist strawberry net sleeve matching by using the deep learning technology, the common practice is that two-dimensional image information of strawberries is acquired through image acquisition equipment, and then, a deep learning model is utilized to extract characteristics of partial shapes or sizes of the strawberries and is used for net sleeve matching, but the scheme still has the defects that the three-dimensional shapes of the strawberries cannot be accurately represented only by extracting single or small number of dimensional characteristics such as fruit diameters and the like based on the two-dimensional images, the complex and various fruit shape differences of the strawberries are difficult to adapt, the adaptation degree still needs to be assisted and improved by depending on earlier sorting and classifying links, the problems of sorting damage, high cost and difficulty in adapting special-shaped fruits are not solved, and the packaging protection effect is difficult to be fully ensured. Disclosure of Invention In view of the above, the invention aims to provide a web-type rule dynamic matching method based on multi-dimensional strawberry size distribution, so as to solve the problem of poor packaging protection effect caused by difficulty in adapting to strawberry shape difference in the traditional packaging method. A web rule dynamic matching method based on multi-dimensional strawberry size distribution comprises the following steps: a1, acquiring strawberry double-view images and strawberry weight data, and respectively preprocessing to obtain preprocessed strawberry double-view images and preprocessed strawberry weight data; a2, respectively extracting initial morphological characteristics for each view according to the preprocessed strawberry double-view image, and then carrying out double-view characteristic fusion to obtain double-view fusion characteristics; A3, inputting the double-view fusion features into a fruit-shaped contour enhancement network, and enhancing and optimizing the contour features to obtain contour enhancement features; a4, extracting three-dimensional morphological characteristics of the strawberries according to the contour enhancement characteristics; A5, extracting weight characteristics of the strawberries according to the pretreated weight data of the strawberries, and combining t