An End-to-End Steel Surface Classification Approach Based on EDCGAN and MobileNet V2.

In the production process of steel products, it is very important to find defects, which can not only reduce the failure rate of industrial production but also can reduce economic losses. All deep learning-based methods need many labeled samples for training. However, in the industrial field, there is a lack of sufficient training samples, especially in steel surface defects. It is almost impossible to collect enough samples that can be used for training. To solve this kind of problem, different from traditional data enhancement methods, this paper constructed a data enhancement model dependent on GAN, using our designed EDCGAN to generate abundant samples that can be used for training. Finally, we mixed different proportions of the generated samples with the original samples and tested them through the MobileNet V2 classification model. The test results showed that if we added the samples generated by EDCGAN to the original samples, the classification results would gradually improve. When the ratio reaches 80%, the overall classification result reaches the highest, achieving an accuracy rate of more than 99%. The experimental process proves the effectiveness of this method and can improve the quality of steel processing.

Fired bullet signature correlation using the finite ridgelet transform (FRIT) and the gray level co-occurrence matrix (GLCM) methods.

When a bullet is fired from a barrel, micro striation marks caused by the sliding motion of the bullet through the rifled barrel are one of the foremost factors in automated ballistic identification. This paper focuses on 3D topography images of land engraved areas (LEA) and proposes a bullet identification method incorporating the finite ridgelet transform (FRIT) and gray level co-occurrence matrix (GLCM) algorithms. The FRIT extracts the striation marks from the 3D micro image and the GLCM generates a linearly weighted weight corresponding to the texture features for 2D average profile calculation. The entire striation marks image is divided into several cells and a cell with valid correlation areas is assigned a large weight, but the one with invalid correlation areas is assigned a small weight along the vertical direction. The visible results show that the valid correlation areas are effectively identified and the negative effects of invalid correlation areas are suppressed. Tests were performed on a control set and an unknown set, giving a total of 35 bullet samples fired from pistols with 10 consecutively manufactured slides. The results included no false identifications or false exclusions and a clear separation between the matching index of the matching and non-matching LEA profiles, demonstrating excellent performance in striation mark capture and valid correlation areas extraction of FRIT and GLCM algorithms. The proposed method is capable of correctly matching toolmarked surfaces to the barrel used.

Convergence-improved congruent matching cells (CMC) method for firing pin impression comparison.

A firing pin impression is usually concave in shape with a small textured area, which makes it difficult to perform automated algorithm-based comparison. The congruent matching cells (CMC) method was invented for accurate breech face impression comparison, in which a reference impression is divided into correlation cells. Each cell is registered to a cell-sized area of the comparison impression that has maximum similarity in surface topography. Four parameters are used to quantify the congruent matching pattern of the registration position and orientation. This paper aims to further develop the cell-division-matching method based on a convergence feature and to develop practical convergence-improved algorithms for firing pin impression comparison. The convergence feature refers to the tendency of the x-y registration positions of correlated cell pairs to converge at the correct registration angle when comparing same-source samples at different orientations. The areal Gaussian filter is employed to extract high-frequency micro-features; the least-squares matching method is used to improve each cross-correlation precision and reach convergence in the registration positions of correlated cell pairs; and a density-based clustering algorithm is introduced to collect the registration positions of dense cell pairs relative to a virtual common center and to remove outliers. Improvements are achieved in the reliability and accuracy of the number of congruent matching cell pairs (CMCs) collected, which represents the quantification of the degree of pairwise impression similarity. Experiments in this report used 40 firing pin impression samples on cartridge cases fired from 10 pistols. The results included no false identifications or false exclusions.